docs: update README with VAD, new threshold option, and silence handling

Update README to clarify VAD mode behavior: provisional caption only after Silero detects speech, Parakeet idle during silence. Document new --vad-threshold option to control Silero speech confidence. Add troubleshooting tip for music/noise false detections.

.gitea/workflows/release.yml

@@ -27,7 +27,7 @@ jobs:
         run: >-
           dnf install -y
           bzip2 curl gcc gcc-c++ git gtk4-devel gtk4-layer-shell-devel
-          jq libX11-devel libXfixes-devel nodejs24 nodejs24-bin pkgconf-pkg-config
+          jq libX11-devel libXfixes-devel nodejs24 nodejs24-bin patchelf pkgconf-pkg-config
           pulseaudio-utils ShellCheck tar gzip binutils
 
       - name: Check out repository
@@ -63,7 +63,7 @@ jobs:
         run: >-
           dnf install -y
           bzip2 curl gcc gcc-c++ git gtk4-devel gtk4-layer-shell-devel
-          jq libX11-devel libXfixes-devel nodejs24 nodejs24-bin pkgconf-pkg-config
+          jq libX11-devel libXfixes-devel nodejs24 nodejs24-bin patchelf pkgconf-pkg-config
           pulseaudio-utils ShellCheck tar gzip binutils
 
       - name: Check out repository and tags
@@ -95,7 +95,7 @@ jobs:
           previous_tag="$(git describe --tags --abbrev=0 "${GITHUB_REF_NAME}^" --match 'v[0-9]*' 2>/dev/null || true)"
           {
             printf 'Automated Captioneer release for `%s`.\n\n' "${GITHUB_REF_NAME}"
-            printf 'The Linux x86_64 archive contains both Captioneer commands and their Sherpa/ONNX native libraries. Models are downloaded separately with `scripts/download-models.sh`.\n'
+            printf 'The Linux x86_64 archive contains both Captioneer commands and a private desktop runtime with GTK4, GTK4 Layer Shell, and Sherpa/ONNX libraries. Models are downloaded separately with `scripts/download-models.sh`.\n'
             if [[ -n "$previous_tag" ]]; then
               printf '\nChanges since `%s`:\n\n' "$previous_tag"
               git log --no-merges --pretty='- %s (`%h`)' "${previous_tag}..${GITHUB_REF_NAME}"

README.md

@@ -11,7 +11,7 @@ Captioneer provides two commands:
 
 ## Setup
 
-The terminal command requires Linux x86_64, Go 1.26.2 or newer, `pactl`, and `parec`. The desktop command additionally needs GTK4 and GTK4 Layer Shell development/runtime libraries. PipeWire works through its PulseAudio compatibility layer.
+The terminal command requires Linux x86_64, Go 1.26.2 or newer, `pactl`, and `parec`. Building the desktop command from source additionally needs GTK4 and GTK4 Layer Shell development libraries. Published release archives carry the desktop runtime libraries. PipeWire works through its PulseAudio compatibility layer.
 
 Download the large recognition models once:
 
@@ -23,7 +23,7 @@ This creates `models/parakeet-tdt-v2/` and `models/silero_vad_v5.onnx`. They are
 
 ## Terminal captions
 
-VAD mode is recommended for normal use. It redraws a provisional caption while speech is active and commits a timestamped caption after the utterance ends:
+VAD mode is recommended for normal use. It redraws a provisional caption only after Silero detects speech and commits a timestamped caption after the utterance ends:
 
 ```bash
 go run ./src/cmd/captioneer --mode=vad
@@ -50,6 +50,14 @@ Final captions use stdout; diagnostics and overload warnings use stderr. This ma
 
 Press <kbd>Ctrl</kbd>+<kbd>C</kbd> to stop. Captioneer stops `parec` and flushes pending audio before exiting.
 
+Sherpa recognition, Silero VAD, and audio capture run in an isolated worker process. The terminal or GTK process stays responsive if native inference crashes or hangs: Captioneer restarts the worker automatically after a failure, and Ctrl+C force-kills only that worker if it cannot flush within the configured shutdown timeout. Send `SIGHUP` to the parent Captioneer PID to hot-reload the models without closing the terminal or overlay:
+
+```bash
+kill -HUP <captioneer-parent-pid>
+```
+
+An automatic or manual restart begins a fresh capture/transcription session; audio being processed by the failed worker cannot be recovered.
+
 ## Desktop overlay
 
 The simplest Bazzite development command is:
@@ -66,6 +74,8 @@ Use `--output=both` to keep terminal output as well:
 
 In VAD mode, a 500 ms pause finalizes the current speech segment. Captioneer splits that transcript again at sentence punctuation and at word-safe boundaries of at most 64 characters, so continuous speech does not become one enormous paragraph. The live provisional view keeps the latest three digestible chunks and grows vertically without ellipsizing them. Completed chunks remain visible for at least three seconds, receive extra time based on their word count, and gradually fade from white to a restrained gray as they age. When several chunks finalize together, their removal is staggered in reading order instead of deleting the whole block at once. Up to six completed chunks remain above the provisional text; if a seventh arrives, the oldest is removed. Blank recognition results never erase completed text before its reading lifetime expires.
 
+The expensive Parakeet recognizer stays idle during silence and audio that Silero does not classify as speech. A small rolling preview window also prevents long continuous audio from making each refresh progressively slower. Silero itself continues receiving the live stream because it needs the quiet frames to detect when speech ends and when it starts again.
+
 Text is left-aligned inside the centered, fixed-width dark bubble and wraps at its configured width instead of extending off-screen. Pointer input passes through the window by default.
 
 To build outside the helper script, install GTK4 and GTK4 Layer Shell development packages, then run:
@@ -97,7 +107,11 @@ Layer Shell and X11 hints are requests, not universal guarantees. Same-layer win
 | `--chunk-duration=1s` | `1s` | Fixed chunk length or VAD preview refresh interval. Accepts Go durations such as `500ms` or `2s`. |
 | `--models-dir=models` | `models` | Directory containing `parakeet-tdt-v2/` and `silero_vad_v5.onnx`. |
 | `--threads=2` | `2` | CPU threads used by recognition and VAD. |
-| `--preview-threshold-dbfs=-45` | `-45` | RMS threshold that starts a provisional VAD caption. Raise it to reject background audio; lower it for quiet speech. |
+| `--preview-threshold-dbfs=-45` | `-45` | RMS gate below which recognition stays idle. Raise it to reject quiet background noise; lower it for quiet speech. In fixed mode, gated chunks are not sent to Parakeet. |
+| `--vad-threshold=0.5` | `0.5` | Silero speech confidence from `0` to `1`. Raise it to reject more music/noise; lower it if real speech is missed. Applies to VAD mode. |
+| `--model-auto-restart=true` | `true` | Restart the isolated model worker after a crash, capture failure, or timeout. Set `false` to report the failure and exit instead. |
+| `--model-timeout=30s` | `30s` | Maximum model startup, continuously busy processing time, or missing-heartbeat period before the worker is considered hung. `0s` disables hang detection. |
+| `--model-shutdown-timeout=2s` | `2s` | Time allowed for pending-audio flush and model cleanup after Ctrl+C or reload before the worker is force-killed. |
 
 ### Desktop-only options
 
@@ -127,6 +141,10 @@ Examples:
 # Fixed two-second captions on a named monitor
 ./scripts/distrobox/run.sh --mode=fixed --chunk-duration=2s \
   --output=overlay --overlay-monitor=DP-1
+
+# Be more conservative when instrumental music causes false speech detections
+./scripts/distrobox/run.sh --mode=vad --output=overlay \
+  --vad-threshold=0.65 --preview-threshold-dbfs=-40
 ```
 
 ## Bazzite development with Distrobox
@@ -162,14 +180,16 @@ Captioneer itself has no environment-based application configuration; use the co
 - **Missing model files:** run `./scripts/download-models.sh` or correct `--models-dir`.
 - **`pactl` or `parec` missing:** install PulseAudio utilities, or use the Distrobox scripts.
 - **No captions:** confirm the current sink with `pactl get-default-sink`; Captioneer listens to `<sink>.monitor`.
+- **Music/noise produces captions:** use VAD mode and try `--vad-threshold=0.65`. Increase it in small steps if false detections continue. `--preview-threshold-dbfs=-40` can also reject quiet background audio, but an RMS gate cannot distinguish loud music from speech. Singing may still be treated as speech.
 - **Processing overload warning:** inference fell behind and the oldest 100 ms packet was dropped. Try more `--threads` or a longer `--chunk-duration`.
 - **Overlay is not always above other windows:** read the backend warning. On Wayland this usually means the compositor lacks Layer Shell; add a compositor rule for the `Captioneer Overlay` window if supported.
 - **Overlay initialization fails:** `--output=overlay` exits with the error. `--output=both` warns and continues terminal-only.
-- **Desktop binary misses shared libraries on Bazzite:** use `scripts/distrobox/run.sh`; container-built GTK binaries are not self-contained.
+- **A locally built desktop binary misses shared libraries:** use `scripts/distrobox/run.sh` or install the matching GTK4 and GTK4 Layer Shell runtime packages. The desktop command in a published release carries these application libraries in its private `lib/` directory; keep the extracted directory together.
+- **The model worker repeatedly restarts:** inspect the warning on stderr. Slow machines may need a larger `--model-timeout`; use `--model-auto-restart=false` when diagnosing a persistent crash so Captioneer exits on the first failure.
 
 ## Development
 
-The transcription pipeline emits presentation-neutral events. `output/terminal` and `output/overlay` consume those events independently; GTK stays on its main OS thread while capture and recognition run on a worker.
+The transcription pipeline emits presentation-neutral events. `output/terminal` and `output/overlay` consume those events independently; GTK stays on its main OS thread while capture, VAD, and recognition run in a supervised child process. Caption events and health messages cross a private pipe, and the child is placed in its own process group so its `parec` process can be stopped or replaced with it.
 
 ```text
 src/app/                 capture/model/process lifecycle
@@ -212,12 +232,12 @@ git tag v1.3.519
 git push origin v1.3.519
 ```
 
-The release contains a checksummed Linux x86_64 archive with `captioneer`, `captioneer-desktop`, the required Sherpa/ONNX shared libraries, the model download script, the README, and the license. The desktop command still requires GTK4 and GTK4 Layer Shell runtime libraries on the destination system.
+The release contains a checksummed Linux x86_64 archive with `captioneer`, `captioneer-desktop`, and a private runtime containing GTK4, GTK4 Layer Shell, Sherpa/ONNX, and their linked non-glibc dependencies. Extract the whole archive and run `./captioneer-desktop`; do not move the executable away from its adjacent `lib/` directory. Models, `pactl`, `parec`, host fonts, the display server, graphics drivers, and glibc remain host-provided.
 
 The release workflow uses Gitea's built-in `GITEA_TOKEN`; no personal token is required. In **Settings → Actions → General**, allow the workflow token `Read` access to code and `Write` access to releases. Gitea clamps workflow permissions to those repository settings, so a more restrictive maximum produces an actionable HTTP 403 during publishing.
 
 ## License
 
-Captioneer is licensed under the [Apache License 2.0](LICENSE). Sherpa-Onnx is Apache-2.0 licensed, GTK4 is LGPL-2.1-or-later, and GTK4 Layer Shell is MIT licensed. Captioneer links to the system GTK libraries rather than copying their source into the project.
+Captioneer is licensed under the [Apache License 2.0](LICENSE). Sherpa-Onnx is Apache-2.0 licensed, GTK4 is LGPL-2.1-or-later, and GTK4 Layer Shell is MIT licensed. Release archives dynamically link and redistribute the required shared libraries; `RUNTIME-LIBRARIES.tsv` records their source packages and `third-party-licenses/` contains the installed license notices supplied by those packages. Local source builds continue to link against the system libraries.
 
 Version 1 intentionally has no animation, persisted caption history, runtime dragging, background-free mode, capture exclusion, Flatpak, desktop entry, or autostart integration.

scripts/ci/package-release.sh

@@ -4,6 +4,13 @@ set -Eeuo pipefail
 repo_root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." && pwd -P)"
 cd -- "$repo_root"
 
+for command in go ldd patchelf readelf rpm; do
+  command -v "$command" >/dev/null 2>&1 || {
+    printf 'captioneer: %s is required to build a release archive\n' "$command" >&2
+    exit 1
+  }
+done
+
 tag=${1:-}
 if [[ ! "$tag" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
   printf 'captioneer: release tag must match vMAJOR.MINOR.BUG, got %q\n' "$tag" >&2
@@ -17,29 +24,189 @@ package_dir="dist/$package_name"
 
 rm -rf -- "$package_dir"
 rm -f -- "dist/$archive_name" "dist/$archive_name.sha256"
-mkdir -p -- "$package_dir/lib" "$package_dir/scripts"
+mkdir -p -- \
+  "$package_dir/lib" \
+  "$package_dir/scripts" \
+  "$package_dir/third-party-licenses"
 
 # The Sherpa Go module embeds an absolute build-machine RPATH. Add a portable
-# fallback so the bundled native libraries resolve beside the extracted tools.
+# build-time fallback; the final paths are normalized after dependency copying.
 # shellcheck disable=SC2016 # $ORIGIN must remain literal for the ELF loader.
-export CGO_LDFLAGS='-Wl,-rpath,$ORIGIN/lib'
+CGO_LDFLAGS='-Wl,-rpath,$ORIGIN/lib' \
-go build -trimpath -o "$package_dir/captioneer" ./src/cmd/captioneer
+  go build -trimpath -o "$package_dir/captioneer" ./src/cmd/captioneer
-go build -trimpath -tags gtk -o "$package_dir/captioneer-desktop" ./src/cmd/captioneer-desktop
+# shellcheck disable=SC2016 # $ORIGIN must remain literal for the ELF loader.
+CGO_LDFLAGS='-Wl,-rpath,$ORIGIN/lib' \
+  go build -trimpath -tags gtk \
+    -o "$package_dir/captioneer-desktop" ./src/cmd/captioneer-desktop
 
 sherpa_dir="$(go list -m -f '{{.Dir}}' github.com/k2-fsa/sherpa-onnx-go-linux)"
 sherpa_lib_dir="$sherpa_dir/lib/x86_64-unknown-linux-gnu"
-cp -- "$sherpa_lib_dir/libsherpa-onnx-c-api.so" "$package_dir/lib/"
+
-cp -- "$sherpa_lib_dir/libonnxruntime.so" "$package_dir/lib/"
+declare -A bundled_libraries=()
+declare -A copied_license_packages=()
+manifest_entries="$(mktemp)"
+trap 'rm -f -- "$manifest_entries"' EXIT
+
+is_host_runtime_library() {
+  case "$1" in
+    ld-linux*.so* | libc.so.* | libm.so.* | libmvec.so.* | libdl.so.* | \
+      libpthread.so.* | librt.so.* | libresolv.so.* | libanl.so.* | \
+      libutil.so.* | libnss_*.so.* | libthread_db.so.* | libBrokenLocale.so.*)
+      return 0
+      ;;
+  esac
+  return 1
+}
+
+copy_package_licenses() {
+  local package=$1
+  local license_path relative_path destination
+
+  [[ -z "${copied_license_packages[$package]:-}" ]] || return 0
+  copied_license_packages["$package"]=1
+
+  while IFS= read -r license_path; do
+    [[ -f "$license_path" && "$license_path" == /usr/share/licenses/* ]] || continue
+    relative_path="${license_path#/usr/share/licenses/}"
+    destination="$package_dir/third-party-licenses/$relative_path"
+    mkdir -p -- "$(dirname -- "$destination")"
+    cp -L -- "$license_path" "$destination"
+  done < <(rpm -ql "$package" 2>/dev/null || true)
+}
+
+bundle_library() {
+  local source=$1
+  local provider=${2:-}
+  local library_name resolved_source destination package
+
+  library_name="$(basename -- "$source")"
+  is_host_runtime_library "$library_name" && return
+
+  resolved_source="$(readlink -f -- "$source")"
+  [[ -f "$resolved_source" ]] || {
+    printf 'captioneer: runtime library not found: %s\n' "$source" >&2
+    exit 1
+  }
+
+  if [[ -n "${bundled_libraries[$library_name]:-}" ]]; then
+    if ! cmp -s -- "$resolved_source" "${bundled_libraries[$library_name]}"; then
+      printf 'captioneer: conflicting runtime libraries named %s\n' "$library_name" >&2
+      exit 1
+    fi
+    return
+  fi
+
+  destination="$package_dir/lib/$library_name"
+  cp -L -- "$resolved_source" "$destination"
+  bundled_libraries["$library_name"]="$resolved_source"
+
+  if [[ -z "$provider" ]] && command -v rpm >/dev/null 2>&1; then
+    package="$(rpm -qf --queryformat '%{NAME}\n' "$resolved_source" 2>/dev/null | head -n 1 || true)"
+    if [[ -n "$package" ]]; then
+      provider="$(rpm -q --queryformat '%{NAME}-%{VERSION}-%{RELEASE}.%{ARCH}' "$package")"
+      copy_package_licenses "$package"
+    fi
+  fi
+  provider=${provider:-unidentified-system-package}
+  printf '%s\t%s\n' "$library_name" "$provider" >> "$manifest_entries"
+}
+
+bundle_library "$sherpa_lib_dir/libsherpa-onnx-c-api.so" \
+  'github.com/k2-fsa/sherpa-onnx-go-linux'
+bundle_library "$sherpa_lib_dir/libonnxruntime.so" \
+  'github.com/k2-fsa/sherpa-onnx-go-linux'
+
+# ldd reports the complete linked dependency closure. Copy everything needed by
+# either command except glibc and its loader, which must match the host system.
+for executable in \
+  "$package_dir/captioneer" \
+  "$package_dir/captioneer-desktop"; do
+  ldd_output="$(ldd "$executable")"
+  if grep -Fq 'not found' <<< "$ldd_output"; then
+    printf 'captioneer: unresolved release dependencies for %s:\n%s\n' \
+      "$executable" "$ldd_output" >&2
+    exit 1
+  fi
+
+  while IFS= read -r library_path; do
+    [[ -n "$library_path" ]] || continue
+    if [[ "$library_path" == "$repo_root/$package_dir/lib/"* ]]; then
+      continue
+    fi
+    bundle_library "$library_path"
+  done < <(
+    awk '
+      $2 == "=>" && $3 ~ /^\// { print $3 }
+      $1 ~ /^\// { print $1 }
+    ' <<< "$ldd_output"
+  )
+done
+
+{
+  printf 'Library\tProvider\n'
+  sort -u -- "$manifest_entries"
+} > "$package_dir/RUNTIME-LIBRARIES.tsv"
+
+# Give every object a location-relative path. Patching each library is required
+# because an upstream RUNPATH (for example, Tinysparql's Fedora path) otherwise
+# prevents the executable's path from resolving that library's own children.
+# This avoids LD_LIBRARY_PATH, which would leak into spawned pactl/parec tools.
+for library in "$package_dir"/lib/*; do
+  chmod u+w -- "$library"
+  # shellcheck disable=SC2016 # $ORIGIN must remain literal for the ELF loader.
+  patchelf --set-rpath '$ORIGIN' "$library"
+  chmod 0755 -- "$library"
+done
+for executable in captioneer captioneer-desktop; do
+  # shellcheck disable=SC2016 # $ORIGIN must remain literal for the ELF loader.
+  patchelf --set-rpath '$ORIGIN/lib' "$package_dir/$executable"
+done
+
 cp -- README.md LICENSE "$package_dir/"
 cp -- scripts/download-models.sh "$package_dir/scripts/"
 
 for executable in captioneer captioneer-desktop; do
+  dynamic_section="$(readelf -d "$package_dir/$executable")"
   # shellcheck disable=SC2016 # Verify the literal ELF-loader token.
-  if ! readelf -d "$package_dir/$executable" | grep -Fq '$ORIGIN/lib'; then
+  if [[ "$dynamic_section" != *'$ORIGIN/lib'* ]]; then
-    printf 'captioneer: %s is missing the portable native-library RPATH\n' "$executable" >&2
+    printf 'captioneer: %s is missing its private-runtime RUNPATH\n' \
+      "$executable" >&2
     exit 1
   fi
 done
+[[ -f "$package_dir/lib/libgtk4-layer-shell.so.0" ]] || {
+  printf 'captioneer: GTK4 Layer Shell was not included in the release runtime\n' >&2
+  exit 1
+}
+
+for library in "$package_dir"/lib/*; do
+  # shellcheck disable=SC2016 # Verify the literal ELF-loader token.
+  [[ "$(patchelf --print-rpath "$library")" == '$ORIGIN' ]] || {
+    printf 'captioneer: %s is missing its private dependency RUNPATH\n' \
+      "$(basename -- "$library")" >&2
+    exit 1
+  }
+done
+
+# Verify the archive itself contains every linked non-glibc dependency instead
+# of allowing the build container's installed libraries to mask an omission.
+for object in \
+  "$package_dir/captioneer" \
+  "$package_dir/captioneer-desktop" \
+  "$package_dir"/lib/*; do
+  while IFS= read -r needed_library; do
+    is_host_runtime_library "$needed_library" && continue
+    [[ -f "$package_dir/lib/$needed_library" ]] || {
+      printf 'captioneer: %s requires unbundled library %s\n' \
+        "$(basename -- "$object")" "$needed_library" >&2
+      exit 1
+    }
+  done < <(patchelf --print-needed "$object")
+done
+
+# --help exits before display/model initialization while still exercising the
+# dynamic loader through the same executable release users invoke.
+"$package_dir/captioneer-desktop" --help >/dev/null 2>&1
 
 tar -C dist -czf "dist/$archive_name" "$package_name"
 (

src/app/run.go

@@ -14,15 +14,28 @@ import (
 	"tea.chunkbyte.com/kato/captioneer/src/output"
 )
 
-func Run(ctx context.Context, settings config.Settings, sink output.Sink, diagnostics io.Writer) error {
+type directHooks struct {
-	if err := capture.ValidatePrograms(); err != nil {
+	ready func() error
+	busy  func() error
+	idle  func() error
+}
+
+func runDirect(ctx context.Context, settings config.Settings, sink output.Sink, diagnostics io.Writer, hooks directHooks) error {
+	if err := hooks.busy(); err != nil {
 		return err
 	}
 	resources, err := models.New(settings)
+	if idleErr := hooks.idle(); err == nil {
+		err = idleErr
+	}
 	if err != nil {
 		return err
 	}
-	defer resources.Close()
+	defer func() {
+		_ = hooks.busy()
+		resources.Close()
+		_ = hooks.idle()
+	}()
 
 	ctx, cancel := context.WithCancel(ctx)
 	defer cancel()
@@ -32,24 +45,41 @@ func Run(ctx context.Context, settings config.Settings, sink output.Sink, diagno
 		return fmt.Errorf("find default output monitor: %w", err)
 	}
 	fmt.Fprintf(diagnostics, "Capturing from: %s\n", monitorSource)
-	fmt.Fprintln(diagnostics, "Press Ctrl+C to stop.")
 
 	packets, waitCapture, err := capture.Packets(ctx, monitorSource)
 	if err != nil {
 		return fmt.Errorf("start system-audio capture: %w", err)
 	}
+	if err := hooks.ready(); err != nil {
+		cancel()
+		_ = waitCapture()
+		return err
+	}
 
 	processor := captions.New(settings, resources, sink.Publish)
 	var processErr error
 	for packet := range packets {
-		if err := processor.Accept(packet); err != nil {
+		if err := hooks.busy(); err != nil {
+			processErr = err
+			cancel()
+			break
+		}
+		acceptErr := processor.Accept(packet)
+		idleErr := hooks.idle()
+		if acceptErr != nil || idleErr != nil {
+			err := errors.Join(acceptErr, idleErr)
 			processErr = fmt.Errorf("publish caption: %w", err)
 			cancel()
 			break
 		}
 	}
 	if processErr == nil {
+		if err := hooks.busy(); err != nil {
+			processErr = err
+		} else {
 			processErr = processor.Flush()
+			processErr = errors.Join(processErr, hooks.idle())
+		}
 	}
 	captureErr := waitCapture()
 	if captureErr != nil && ctx.Err() == nil {

src/app/supervisor.go

@@ -0,0 +1,361 @@
+package app
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"os/signal"
+	"strings"
+	"syscall"
+	"time"
+
+	"tea.chunkbyte.com/kato/captioneer/src/capture"
+	"tea.chunkbyte.com/kato/captioneer/src/config"
+	"tea.chunkbyte.com/kato/captioneer/src/models"
+	"tea.chunkbyte.com/kato/captioneer/src/output"
+)
+
+const modelRestartDelay = time.Second
+
+type attemptResult uint8
+
+const (
+	attemptStopped attemptResult = iota + 1
+	attemptReload
+	attemptFailed
+	attemptFatal
+)
+
+type workerHealth struct {
+	startedAt time.Time
+	lastSeen  time.Time
+	ready     bool
+	busySince time.Time
+}
+
+func (h *workerHealth) observe(messageType string, now time.Time) {
+	h.lastSeen = now
+	switch messageType {
+	case messageReady:
+		h.ready = true
+		h.busySince = time.Time{}
+	case messageBusy:
+		if h.busySince.IsZero() {
+			h.busySince = now
+		}
+	case messageIdle:
+		h.busySince = time.Time{}
+	}
+}
+
+func (h workerHealth) timeoutError(now time.Time, timeout time.Duration) error {
+	if timeout == 0 {
+		return nil
+	}
+	if !h.ready && now.Sub(h.startedAt) >= timeout {
+		return fmt.Errorf("model worker did not become ready within %s", timeout)
+	}
+	if !h.busySince.IsZero() && now.Sub(h.busySince) >= timeout {
+		return fmt.Errorf("model operation exceeded %s", timeout)
+	}
+	if h.ready && now.Sub(h.lastSeen) >= timeout {
+		return fmt.Errorf("model worker heartbeat was silent for %s", timeout)
+	}
+	return nil
+}
+
+// Run supervises an isolated process containing capture, VAD, and recognition.
+// Presentation stays in this process, so the native model can be replaced
+// without restarting the terminal or GTK main loop.
+func Run(ctx context.Context, settings config.Settings, sink output.Sink, diagnostics io.Writer) error {
+	if err := capture.ValidatePrograms(); err != nil {
+		return err
+	}
+	if err := models.Validate(settings); err != nil {
+		return err
+	}
+
+	fmt.Fprintln(diagnostics, "Press Ctrl+C to stop. Send SIGHUP to hot-reload the model worker.")
+	reload := make(chan os.Signal, 1)
+	signal.Notify(reload, syscall.SIGHUP)
+	defer signal.Stop(reload)
+
+	for {
+		result, err := runWorkerAttempt(ctx, settings, sink, diagnostics, reload)
+		if ctx.Err() != nil {
+			return nil
+		}
+		switch result {
+		case attemptStopped:
+			return nil
+		case attemptFatal:
+			return err
+		case attemptReload:
+			fmt.Fprintln(diagnostics, "Reloading model worker.")
+			continue
+		case attemptFailed:
+			if !settings.ModelAutoRestart {
+				return err
+			}
+			fmt.Fprintf(diagnostics, "warning: %v; restarting model worker in %s\n", err, modelRestartDelay)
+		}
+
+		timer := time.NewTimer(modelRestartDelay)
+		select {
+		case <-ctx.Done():
+			if !timer.Stop() {
+				<-timer.C
+			}
+			return nil
+		case <-reload:
+			if !timer.Stop() {
+				<-timer.C
+			}
+			fmt.Fprintln(diagnostics, "Reloading model worker.")
+		case <-timer.C:
+		}
+	}
+}
+
+func runWorkerAttempt(
+	ctx context.Context,
+	settings config.Settings,
+	sink output.Sink,
+	diagnostics io.Writer,
+	reload <-chan os.Signal,
+) (attemptResult, error) {
+	cmd, protocol, err := startModelWorker(settings, diagnostics)
+	if err != nil {
+		return attemptFailed, fmt.Errorf("start model worker: %w", err)
+	}
+	defer protocol.Close()
+
+	messages := make(chan workerMessage)
+	protocolDone := make(chan error, 1)
+	go readWorkerMessages(protocol, messages, protocolDone)
+
+	wait := make(chan error, 1)
+	go func() { wait <- cmd.Wait() }()
+
+	checkInterval := 250 * time.Millisecond
+	if settings.ModelTimeout > 0 && settings.ModelTimeout < checkInterval {
+		checkInterval = settings.ModelTimeout / 2
+		if checkInterval <= 0 {
+			checkInterval = time.Millisecond
+		}
+	}
+	ticker := time.NewTicker(checkInterval)
+	defer ticker.Stop()
+
+	startedAt := time.Now()
+	health := workerHealth{startedAt: startedAt, lastSeen: startedAt}
+	messageChannel := (<-chan workerMessage)(messages)
+	waitChannel := (<-chan error)(wait)
+	contextDone := ctx.Done()
+	var stopTimer *time.Timer
+	var stopDeadline <-chan time.Time
+	var stopping bool
+	var desiredResult attemptResult
+	var desiredErr error
+	var processDone bool
+	var processErr error
+	var protocolClosed bool
+	var protocolErr error
+	var workerFailure string
+
+	beginStop := func(result attemptResult, stopErr error) {
+		if stopping {
+			if result == attemptFatal {
+				desiredResult = result
+				desiredErr = errors.Join(desiredErr, stopErr)
+			}
+			return
+		}
+		stopping = true
+		desiredResult = result
+		desiredErr = stopErr
+		if err := signalWorkerGroup(cmd, syscall.SIGTERM); err != nil {
+			desiredErr = errors.Join(desiredErr, err)
+		}
+		stopTimer = time.NewTimer(settings.ModelShutdownTimeout)
+		stopDeadline = stopTimer.C
+	}
+
+	for {
+		if processDone && protocolClosed {
+			if stopTimer != nil && !stopTimer.Stop() {
+				select {
+				case <-stopTimer.C:
+				default:
+				}
+			}
+			if stopping {
+				return desiredResult, desiredErr
+			}
+			// The worker may have crashed before its parec child. A process group
+			// remains signalable after its leader exits, so clean up any survivors
+			// before starting a replacement worker.
+			cleanupErr := signalWorkerGroup(cmd, syscall.SIGKILL)
+			if protocolErr != nil {
+				return attemptFailed, errors.Join(fmt.Errorf("model worker protocol: %w", protocolErr), cleanupErr)
+			}
+			if workerFailure != "" {
+				return attemptFailed, errors.Join(errors.New(workerFailure), cleanupErr)
+			}
+			if processErr != nil {
+				return attemptFailed, errors.Join(fmt.Errorf("model worker exited: %w", processErr), cleanupErr)
+			}
+			return attemptFailed, errors.Join(errors.New("model worker exited unexpectedly"), cleanupErr)
+		}
+
+		select {
+		case <-contextDone:
+			contextDone = nil
+			beginStop(attemptStopped, nil)
+
+		case <-reload:
+			beginStop(attemptReload, nil)
+
+		case message, ok := <-messageChannel:
+			if !ok {
+				messageChannel = nil
+				protocolClosed = true
+				protocolErr = <-protocolDone
+				if !processDone && !stopping {
+					beginStop(attemptFailed, errors.New("model worker protocol closed unexpectedly"))
+				}
+				continue
+			}
+
+			now := time.Now()
+			health.observe(message.Type, now)
+			switch message.Type {
+			case messageReady:
+				fmt.Fprintln(diagnostics, "Model worker ready.")
+			case messageBusy, messageIdle:
+			case messageHeartbeat:
+			case messageCaption:
+				if message.Event == nil {
+					beginStop(attemptFailed, errors.New("model worker sent an empty caption event"))
+					continue
+				}
+				if err := sink.Publish(*message.Event); err != nil {
+					beginStop(attemptFatal, fmt.Errorf("publish caption: %w", err))
+				}
+			case messageFailure:
+				workerFailure = message.Error
+			default:
+				beginStop(attemptFailed, fmt.Errorf("unknown model worker message %q", message.Type))
+			}
+
+		case err := <-waitChannel:
+			waitChannel = nil
+			processDone = true
+			processErr = err
+
+		case now := <-ticker.C:
+			if !stopping {
+				if err := health.timeoutError(now, settings.ModelTimeout); err != nil {
+					beginStop(attemptFailed, err)
+				}
+			}
+
+		case <-stopDeadline:
+			stopDeadline = nil
+			if !processDone {
+				fmt.Fprintf(diagnostics, "warning: model worker did not stop within %s; force-killing it\n", settings.ModelShutdownTimeout)
+				if err := signalWorkerGroup(cmd, syscall.SIGKILL); err != nil {
+					desiredErr = errors.Join(desiredErr, err)
+				}
+			}
+		}
+	}
+}
+
+func startModelWorker(settings config.Settings, diagnostics io.Writer) (*exec.Cmd, *os.File, error) {
+	executable, err := os.Executable()
+	if err != nil {
+		return nil, nil, err
+	}
+	encodedSettings, err := encodeWorkerSettings(settings)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	protocolRead, protocolWrite, err := os.Pipe()
+	if err != nil {
+		return nil, nil, err
+	}
+
+	cmd := exec.Command(executable)
+	cmd.Env = workerEnvironment(encodedSettings)
+	cmd.ExtraFiles = []*os.File{protocolWrite}
+	cmd.Stdout = diagnostics
+	cmd.Stderr = diagnostics
+	cmd.SysProcAttr = &syscall.SysProcAttr{
+		Setpgid:   true,
+		Pdeathsig: syscall.SIGKILL,
+	}
+	if err := cmd.Start(); err != nil {
+		protocolRead.Close()
+		protocolWrite.Close()
+		return nil, nil, err
+	}
+	if err := protocolWrite.Close(); err != nil {
+		_ = signalWorkerGroup(cmd, syscall.SIGKILL)
+		_ = cmd.Wait()
+		protocolRead.Close()
+		return nil, nil, err
+	}
+	return cmd, protocolRead, nil
+}
+
+func workerEnvironment(encodedSettings string) []string {
+	environment := make([]string, 0, len(os.Environ())+2)
+	for _, item := range os.Environ() {
+		if strings.HasPrefix(item, modelWorkerEnvironment+"=") ||
+			strings.HasPrefix(item, modelSettingsEnvironment+"=") {
+			continue
+		}
+		environment = append(environment, item)
+	}
+	return append(environment,
+		modelWorkerEnvironment+"=1",
+		modelSettingsEnvironment+"="+encodedSettings,
+	)
+}
+
+func readWorkerMessages(protocol io.Reader, messages chan<- workerMessage, done chan<- error) {
+	defer close(messages)
+	decoder := json.NewDecoder(protocol)
+	for {
+		var message workerMessage
+		if err := decoder.Decode(&message); err != nil {
+			if errors.Is(err, io.EOF) {
+				done <- nil
+			} else {
+				done <- err
+			}
+			return
+		}
+		messages <- message
+	}
+}
+
+func signalWorkerGroup(cmd *exec.Cmd, signal syscall.Signal) error {
+	if cmd.Process == nil {
+		return nil
+	}
+	err := syscall.Kill(-cmd.Process.Pid, signal)
+	if errors.Is(err, syscall.ESRCH) {
+		return nil
+	}
+	if err != nil {
+		return fmt.Errorf("signal model worker: %w", err)
+	}
+	return nil
+}

src/app/supervisor_test.go

@@ -0,0 +1,67 @@
+package app
+
+import (
+	"strings"
+	"testing"
+	"time"
+
+	"tea.chunkbyte.com/kato/captioneer/src/config"
+)
+
+func TestWorkerHealthDetectsStartupAndOperationTimeouts(t *testing.T) {
+	started := time.Unix(100, 0)
+	health := workerHealth{startedAt: started, lastSeen: started}
+
+	if err := health.timeoutError(started.Add(time.Second), 2*time.Second); err != nil {
+		t.Fatalf("healthy startup reported a timeout: %v", err)
+	}
+	if err := health.timeoutError(started.Add(2*time.Second), 2*time.Second); err == nil ||
+		!strings.Contains(err.Error(), "did not become ready") {
+		t.Fatalf("startup timeout = %v", err)
+	}
+
+	health.observe(messageReady, started.Add(2*time.Second))
+	health.observe(messageBusy, started.Add(3*time.Second))
+	if err := health.timeoutError(started.Add(4*time.Second), 2*time.Second); err != nil {
+		t.Fatalf("healthy operation reported a timeout: %v", err)
+	}
+	if err := health.timeoutError(started.Add(5*time.Second), 2*time.Second); err == nil ||
+		!strings.Contains(err.Error(), "operation exceeded") {
+		t.Fatalf("operation timeout = %v", err)
+	}
+
+	health.observe(messageIdle, started.Add(5*time.Second))
+	health.observe(messageHeartbeat, started.Add(6*time.Second))
+	if err := health.timeoutError(started.Add(7*time.Second), 2*time.Second); err != nil {
+		t.Fatalf("idle worker reported a timeout: %v", err)
+	}
+	if err := health.timeoutError(started.Add(8*time.Second), 2*time.Second); err == nil ||
+		!strings.Contains(err.Error(), "heartbeat was silent") {
+		t.Fatalf("heartbeat timeout = %v", err)
+	}
+}
+
+func TestWorkerSettingsRoundTrip(t *testing.T) {
+	want := config.Settings{
+		Mode:                 config.ModeVAD,
+		ChunkDuration:        time.Second,
+		ModelsDir:            "models with spaces",
+		Threads:              3,
+		PreviewThresholdDB:   -42,
+		VADThreshold:         0.65,
+		ModelAutoRestart:     true,
+		ModelTimeout:         12 * time.Second,
+		ModelShutdownTimeout: 2 * time.Second,
+	}
+	encoded, err := encodeWorkerSettings(want)
+	if err != nil {
+		t.Fatal(err)
+	}
+	got, err := decodeWorkerSettings(encoded)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got != want {
+		t.Fatalf("settings round trip = %#v, want %#v", got, want)
+	}
+}

src/app/worker_protocol.go

@@ -0,0 +1,152 @@
+package app
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/signal"
+	"sync"
+	"syscall"
+	"time"
+
+	"tea.chunkbyte.com/kato/captioneer/src/captions"
+	"tea.chunkbyte.com/kato/captioneer/src/config"
+)
+
+const (
+	modelWorkerEnvironment   = "CAPTIONEER_INTERNAL_MODEL_WORKER"
+	modelSettingsEnvironment = "CAPTIONEER_INTERNAL_MODEL_SETTINGS"
+	modelProtocolFD          = 3
+)
+
+const (
+	messageReady     = "ready"
+	messageBusy      = "busy"
+	messageIdle      = "idle"
+	messageCaption   = "caption"
+	messageFailure   = "failure"
+	messageHeartbeat = "heartbeat"
+)
+
+type workerMessage struct {
+	Type  string          `json:"type"`
+	Event *captions.Event `json:"event,omitempty"`
+	Error string          `json:"error,omitempty"`
+}
+
+type workerReporter struct {
+	mu      sync.Mutex
+	encoder *json.Encoder
+}
+
+func (r *workerReporter) send(message workerMessage) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.encoder.Encode(message)
+}
+
+type protocolSink struct {
+	reporter *workerReporter
+}
+
+func (s protocolSink) Publish(event captions.Event) error {
+	return s.reporter.send(workerMessage{Type: messageCaption, Event: &event})
+}
+
+func (protocolSink) Close() error { return nil }
+
+// IsModelWorker reports whether this process was created as Captioneer's
+// private inference worker. The environment marker is intentionally internal.
+func IsModelWorker() bool {
+	return os.Getenv(modelWorkerEnvironment) == "1"
+}
+
+func encodeWorkerSettings(settings config.Settings) (string, error) {
+	data, err := json.Marshal(settings)
+	if err != nil {
+		return "", err
+	}
+	return base64.RawStdEncoding.EncodeToString(data), nil
+}
+
+func decodeWorkerSettings(value string) (config.Settings, error) {
+	data, err := base64.RawStdEncoding.DecodeString(value)
+	if err != nil {
+		return config.Settings{}, fmt.Errorf("decode model worker settings: %w", err)
+	}
+	var settings config.Settings
+	if err := json.Unmarshal(data, &settings); err != nil {
+		return config.Settings{}, fmt.Errorf("parse model worker settings: %w", err)
+	}
+	if err := settings.Validate(); err != nil {
+		return config.Settings{}, fmt.Errorf("validate model worker settings: %w", err)
+	}
+	return settings, nil
+}
+
+// RunModelWorker owns all Sherpa/VAD resources in the supervised child. It
+// returns a process exit code so command entrypoints can call os.Exit.
+func RunModelWorker(diagnostics io.Writer) int {
+	protocol := os.NewFile(uintptr(modelProtocolFD), "captioneer-model-protocol")
+	if protocol == nil {
+		fmt.Fprintln(diagnostics, "captioneer: model worker protocol is unavailable")
+		return 1
+	}
+	defer protocol.Close()
+	syscall.CloseOnExec(modelProtocolFD)
+
+	settings, err := decodeWorkerSettings(os.Getenv(modelSettingsEnvironment))
+	if err != nil {
+		fmt.Fprintf(diagnostics, "captioneer: %v\n", err)
+		return 1
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
+	defer stop()
+
+	reporter := &workerReporter{encoder: json.NewEncoder(protocol)}
+	hooks := directHooks{
+		ready: func() error { return reporter.send(workerMessage{Type: messageReady}) },
+		busy:  func() error { return reporter.send(workerMessage{Type: messageBusy}) },
+		idle:  func() error { return reporter.send(workerMessage{Type: messageIdle}) },
+	}
+
+	heartbeatInterval := time.Second
+	if settings.ModelTimeout > 0 && settings.ModelTimeout < 4*heartbeatInterval {
+		heartbeatInterval = settings.ModelTimeout / 4
+		if heartbeatInterval <= 0 {
+			heartbeatInterval = time.Millisecond
+		}
+	}
+	heartbeatContext, stopHeartbeat := context.WithCancel(ctx)
+	heartbeatDone := make(chan struct{})
+	go func() {
+		defer close(heartbeatDone)
+		ticker := time.NewTicker(heartbeatInterval)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-heartbeatContext.Done():
+				return
+			case <-ticker.C:
+				if reporter.send(workerMessage{Type: messageHeartbeat}) != nil {
+					return
+				}
+			}
+		}
+	}()
+
+	err = runDirect(ctx, settings, protocolSink{reporter: reporter}, diagnostics, hooks)
+	stopHeartbeat()
+	<-heartbeatDone
+	if err == nil || errors.Is(err, context.Canceled) {
+		return 0
+	}
+	_ = reporter.send(workerMessage{Type: messageFailure, Error: err.Error()})
+	fmt.Fprintf(diagnostics, "captioneer: model worker: %v\n", err)
+	return 1
+}

src/captions/event.go

@@ -20,6 +20,7 @@ type Event struct {
 type Transcriber interface {
 	Decode(samples []float32) string
 	AcceptVAD(samples []float32)
+	SpeechActive() bool
 	FlushVAD()
 	NextSpeechSegment() (start int, samples []float32, ok bool)
 }

src/captions/processor.go

@@ -3,11 +3,18 @@ package captions
 
 import (
 	"strings"
+	"time"
 
 	"tea.chunkbyte.com/kato/captioneer/src/audio"
 	"tea.chunkbyte.com/kato/captioneer/src/config"
 )
 
+// Provisional recognition uses only a recent window. Re-decoding an entire
+// long utterance every refresh grows increasingly expensive, while the UI only
+// presents the latest provisional lines. The final event still decodes the
+// complete VAD segment.
+const maxPreviewDuration = 10 * time.Second
+
 type Processor struct {
 	settings    config.Settings
 	transcriber Transcriber
@@ -18,6 +25,7 @@ type Processor struct {
 	previewActive   bool
 	previewStart    int
 	nextPreviewAt   int
+	previewReceived int
 	totalSamples    int
 }
 
@@ -43,19 +51,26 @@ func (p *Processor) Accept(samples []float32) error {
 func (p *Processor) Flush() error {
 	if p.settings.Mode == config.ModeFixed {
 		if len(p.fixedSamples) > 0 {
-			return p.emitFinal(p.totalSamples, p.totalSamples+len(p.fixedSamples), p.fixedSamples)
+			return p.emitFixedFinal(p.totalSamples, p.totalSamples+len(p.fixedSamples), p.fixedSamples)
 		}
 		return nil
 	}
 	p.transcriber.FlushVAD()
-	return p.drainVAD()
+	if err := p.drainVAD(); err != nil {
+		return err
+	}
+	if p.previewActive {
+		p.resetPreview()
+		return p.emit(Event{Kind: Hide})
+	}
+	return nil
 }
 
 func (p *Processor) acceptFixed(samples []float32) error {
 	p.fixedSamples = append(p.fixedSamples, samples...)
 	chunkSize := audio.DurationSamples(p.settings.ChunkDuration)
 	for len(p.fixedSamples) >= chunkSize {
-		if err := p.emitFinal(p.totalSamples, p.totalSamples+chunkSize, p.fixedSamples[:chunkSize]); err != nil {
+		if err := p.emitFixedFinal(p.totalSamples, p.totalSamples+chunkSize, p.fixedSamples[:chunkSize]); err != nil {
 			return err
 		}
 		p.fixedSamples = p.fixedSamples[chunkSize:]
@@ -68,20 +83,33 @@ func (p *Processor) acceptVAD(samples []float32) error {
 	p.totalSamples += len(samples)
 	p.transcriber.AcceptVAD(samples)
 
-	if p.previewActive || audio.RMSDBFS(samples) >= p.settings.PreviewThresholdDB {
+	speechActive := p.transcriber.SpeechActive()
+	if speechActive && (p.previewActive || audio.RMSDBFS(samples) >= p.settings.PreviewThresholdDB) {
 		if !p.previewActive {
 			p.previewActive = true
 			p.previewStart = p.totalSamples - len(samples)
 		}
 		p.previewSamples = append(p.previewSamples, samples...)
-		if len(p.previewSamples) >= p.nextPreviewAt {
+		p.previewReceived += len(samples)
+		p.trimPreviewWindow()
+		if p.previewReceived >= p.nextPreviewAt {
 			if err := p.emitProvisional(); err != nil {
 				return err
 			}
 			p.nextPreviewAt += audio.DurationSamples(p.settings.ChunkDuration)
 		}
 	}
-	return p.drainVAD()
+	if err := p.drainVAD(); err != nil {
+		return err
+	}
+	// A short VAD false-positive can become active without meeting the minimum
+	// speech duration, so no final segment is produced. Clear its provisional
+	// text instead of leaving a stale caption on screen.
+	if !speechActive && p.previewActive {
+		p.resetPreview()
+		return p.emit(Event{Kind: Hide})
+	}
+	return nil
 }
 
 func (p *Processor) drainVAD() error {
@@ -93,11 +121,26 @@ func (p *Processor) drainVAD() error {
 		if err := p.emitFinal(start, start+len(samples), samples); err != nil {
 			return err
 		}
+		p.resetPreview()
+	}
+}
+
+func (p *Processor) trimPreviewWindow() {
+	maxSamples := audio.DurationSamples(maxPreviewDuration)
+	if len(p.previewSamples) <= maxSamples {
+		return
+	}
+	dropped := len(p.previewSamples) - maxSamples
+	p.previewSamples = p.previewSamples[dropped:]
+	p.previewStart += dropped
+}
+
+func (p *Processor) resetPreview() {
 	p.previewSamples = nil
 	p.previewActive = false
 	p.previewStart = 0
+	p.previewReceived = 0
 	p.nextPreviewAt = audio.DurationSamples(p.settings.ChunkDuration)
-	}
 }
 
 func (p *Processor) emitProvisional() error {
@@ -125,3 +168,10 @@ func (p *Processor) emitFinal(start, end int, samples []float32) error {
 		EndedAt:   audio.SamplesDuration(end),
 	})
 }
+
+func (p *Processor) emitFixedFinal(start, end int, samples []float32) error {
+	if audio.RMSDBFS(samples) < p.settings.PreviewThresholdDB {
+		return p.emit(Event{Kind: Hide})
+	}
+	return p.emitFinal(start, end, samples)
+}

src/captions/processor_test.go

@@ -18,9 +18,13 @@ type fakeTranscriber struct {
 	decodeTexts  []string
 	segments     []fakeSegment
 	flushed      bool
+	speechActive bool
+	decodeSizes  []int
+	acceptedVAD  int
 }
 
-func (f *fakeTranscriber) Decode([]float32) string {
+func (f *fakeTranscriber) Decode(samples []float32) string {
+	f.decodeSizes = append(f.decodeSizes, len(samples))
 	if len(f.decodeTexts) == 0 {
 		return f.text
 	}
@@ -28,7 +32,8 @@ func (f *fakeTranscriber) Decode([]float32) string {
 	f.decodeTexts = f.decodeTexts[1:]
 	return text
 }
-func (f *fakeTranscriber) AcceptVAD([]float32) {}
+func (f *fakeTranscriber) AcceptVAD(samples []float32) { f.acceptedVAD += len(samples) }
+func (f *fakeTranscriber) SpeechActive() bool          { return f.speechActive }
 func (f *fakeTranscriber) FlushVAD()                   { f.flushed = true }
 func (f *fakeTranscriber) NextSpeechSegment() (int, []float32, bool) {
 	if len(f.segments) == 0 {
@@ -42,12 +47,14 @@ func (f *fakeTranscriber) NextSpeechSegment() (int, []float32, bool) {
 func TestFixedModeEmitsFinalCaption(t *testing.T) {
 	transcriber := &fakeTranscriber{text: "hello"}
 	var events []Event
-	processor := New(config.Settings{Mode: config.ModeFixed, ChunkDuration: time.Second}, transcriber, func(event Event) error {
+	processor := New(config.Settings{
+		Mode: config.ModeFixed, ChunkDuration: time.Second, PreviewThresholdDB: -45,
+	}, transcriber, func(event Event) error {
 		events = append(events, event)
 		return nil
 	})
 
-	if err := processor.Accept(make([]float32, audio.SampleRate)); err != nil {
+	if err := processor.Accept(constantSamples(audio.SampleRate, 0.5)); err != nil {
 		t.Fatal(err)
 	}
 	if len(events) != 1 || events[0].Kind != Final || events[0].Text != "hello" {
@@ -61,12 +68,14 @@ func TestFixedModeEmitsFinalCaption(t *testing.T) {
 func TestFixedModeFlushesFinalPartialChunkWithExactTimestamps(t *testing.T) {
 	transcriber := &fakeTranscriber{decodeTexts: []string{"first", "partial"}}
 	var events []Event
-	processor := New(config.Settings{Mode: config.ModeFixed, ChunkDuration: time.Second}, transcriber, func(event Event) error {
+	processor := New(config.Settings{
+		Mode: config.ModeFixed, ChunkDuration: time.Second, PreviewThresholdDB: -45,
+	}, transcriber, func(event Event) error {
 		events = append(events, event)
 		return nil
 	})
 
-	if err := processor.Accept(make([]float32, audio.SampleRate+audio.SampleRate/2)); err != nil {
+	if err := processor.Accept(constantSamples(audio.SampleRate+audio.SampleRate/2, 0.5)); err != nil {
 		t.Fatal(err)
 	}
 	if err := processor.Flush(); err != nil {
@@ -81,7 +90,7 @@ func TestFixedModeFlushesFinalPartialChunkWithExactTimestamps(t *testing.T) {
 }
 
 func TestVADModeEmitsProvisionalThenFinal(t *testing.T) {
-	transcriber := &fakeTranscriber{text: "speech"}
+	transcriber := &fakeTranscriber{text: "speech", speechActive: true}
 	var events []Event
 	processor := New(config.Settings{
 		Mode:               config.ModeVAD,
@@ -92,14 +101,12 @@ func TestVADModeEmitsProvisionalThenFinal(t *testing.T) {
 		return nil
 	})
 
-	loud := make([]float32, audio.SampleRate)
+	loud := constantSamples(audio.SampleRate, 0.5)
-	for i := range loud {
-		loud[i] = 0.5
-	}
 	if err := processor.Accept(loud); err != nil {
 		t.Fatal(err)
 	}
 	transcriber.segments = []fakeSegment{{start: 0, samples: loud}}
+	transcriber.speechActive = false
 	if err := processor.Accept(make([]float32, audio.SamplesPerPacket())); err != nil {
 		t.Fatal(err)
 	}
@@ -112,11 +119,13 @@ func TestVADModeEmitsProvisionalThenFinal(t *testing.T) {
 func TestEmptyFinalEmitsHide(t *testing.T) {
 	transcriber := &fakeTranscriber{text: " \n\t "}
 	var event Event
-	processor := New(config.Settings{Mode: config.ModeFixed, ChunkDuration: time.Second}, transcriber, func(got Event) error {
+	processor := New(config.Settings{
+		Mode: config.ModeFixed, ChunkDuration: time.Second, PreviewThresholdDB: -45,
+	}, transcriber, func(got Event) error {
 		event = got
 		return nil
 	})
-	if err := processor.Accept(make([]float32, audio.SampleRate)); err != nil {
+	if err := processor.Accept(constantSamples(audio.SampleRate, 0.5)); err != nil {
 		t.Fatal(err)
 	}
 	if event.Kind != Hide {
@@ -124,6 +133,101 @@ func TestEmptyFinalEmitsHide(t *testing.T) {
 	}
 }
 
+func TestFixedModeSilenceStaysOutOfRecognizer(t *testing.T) {
+	transcriber := &fakeTranscriber{text: "hallucinated caption"}
+	var events []Event
+	processor := New(config.Settings{
+		Mode:               config.ModeFixed,
+		ChunkDuration:      time.Second,
+		PreviewThresholdDB: -45,
+	}, transcriber, func(event Event) error {
+		events = append(events, event)
+		return nil
+	})
+
+	if err := processor.Accept(make([]float32, audio.SampleRate)); err != nil {
+		t.Fatal(err)
+	}
+	if len(transcriber.decodeSizes) != 0 {
+		t.Fatalf("silent audio reached recognizer: decode sizes = %v", transcriber.decodeSizes)
+	}
+	if len(events) != 1 || events[0].Kind != Hide {
+		t.Fatalf("silent chunk events = %#v, want one Hide", events)
+	}
+}
+
+func TestVADModeDoesNotPreviewLoudNonSpeech(t *testing.T) {
+	transcriber := &fakeTranscriber{text: "music hallucination"}
+	var events []Event
+	processor := New(config.Settings{
+		Mode:               config.ModeVAD,
+		ChunkDuration:      time.Second,
+		PreviewThresholdDB: -45,
+	}, transcriber, func(event Event) error {
+		events = append(events, event)
+		return nil
+	})
+
+	loudMusic := constantSamples(audio.SampleRate, 0.5)
+	if err := processor.Accept(loudMusic); err != nil {
+		t.Fatal(err)
+	}
+	if transcriber.acceptedVAD != len(loudMusic) {
+		t.Fatalf("VAD received %d samples, want %d", transcriber.acceptedVAD, len(loudMusic))
+	}
+	if len(transcriber.decodeSizes) != 0 || len(events) != 0 {
+		t.Fatalf("non-speech reached recognizer: decodes=%v events=%#v", transcriber.decodeSizes, events)
+	}
+}
+
+func TestVADModeClearsRejectedShortSpeechPreview(t *testing.T) {
+	transcriber := &fakeTranscriber{text: "draft", speechActive: true}
+	var events []Event
+	processor := New(config.Settings{
+		Mode:               config.ModeVAD,
+		ChunkDuration:      time.Second,
+		PreviewThresholdDB: -45,
+	}, transcriber, func(event Event) error {
+		events = append(events, event)
+		return nil
+	})
+
+	if err := processor.Accept(constantSamples(audio.SampleRate, 0.5)); err != nil {
+		t.Fatal(err)
+	}
+	transcriber.speechActive = false
+	if err := processor.Accept(make([]float32, audio.SamplesPerPacket())); err != nil {
+		t.Fatal(err)
+	}
+	if len(events) != 2 || events[0].Kind != Provisional || events[1].Kind != Hide {
+		t.Fatalf("events = %#v, want Provisional then Hide", events)
+	}
+}
+
+func TestVADPreviewRecognitionWindowIsBounded(t *testing.T) {
+	transcriber := &fakeTranscriber{text: "draft", speechActive: true}
+	processor := New(config.Settings{
+		Mode:               config.ModeVAD,
+		ChunkDuration:      time.Second,
+		PreviewThresholdDB: -45,
+	}, transcriber, func(Event) error { return nil })
+
+	for range 15 {
+		if err := processor.Accept(constantSamples(audio.SampleRate, 0.5)); err != nil {
+			t.Fatal(err)
+		}
+	}
+	maxSamples := audio.DurationSamples(maxPreviewDuration)
+	for _, size := range transcriber.decodeSizes {
+		if size > maxSamples {
+			t.Fatalf("preview decode used %d samples, limit is %d", size, maxSamples)
+		}
+	}
+	if len(transcriber.decodeSizes) != 15 {
+		t.Fatalf("preview decode count = %d, want 15", len(transcriber.decodeSizes))
+	}
+}
+
 func TestFlushFlushesVAD(t *testing.T) {
 	transcriber := &fakeTranscriber{}
 	processor := New(config.Settings{Mode: config.ModeVAD, ChunkDuration: time.Second}, transcriber, func(Event) error { return nil })
@@ -134,3 +238,11 @@ func TestFlushFlushesVAD(t *testing.T) {
 		t.Fatal("VAD was not flushed")
 	}
 }
+
+func constantSamples(count int, value float32) []float32 {
+	samples := make([]float32, count)
+	for i := range samples {
+		samples[i] = value
+	}
+	return samples
+}

src/capture/pulse.go

@@ -10,6 +10,7 @@ import (
 	"os/exec"
 	"strings"
 	"sync"
+	"syscall"
 
 	"tea.chunkbyte.com/kato/captioneer/src/audio"
 )
@@ -41,6 +42,7 @@ func Packets(ctx context.Context, monitorSource string) (<-chan []float32, func(
 		"--latency-msec=50",
 	)
 	cmd.Stderr = os.Stderr
+	cmd.SysProcAttr = &syscall.SysProcAttr{Pdeathsig: syscall.SIGKILL}
 	stdout, err := cmd.StdoutPipe()
 	if err != nil {
 		return nil, nil, err

src/cmd/captioneer-desktop/main.go

@@ -20,6 +20,10 @@ import (
 )
 
 func main() {
+	if app.IsModelWorker() {
+		os.Exit(app.RunModelWorker(os.Stderr))
+	}
+
 	runtime.LockOSThread()
 	defer runtime.UnlockOSThread()
 

src/cmd/captioneer/main.go

@@ -13,6 +13,10 @@ import (
 )
 
 func main() {
+	if app.IsModelWorker() {
+		os.Exit(app.RunModelWorker(os.Stderr))
+	}
+
 	settings := config.Parse()
 	if err := settings.Validate(); err != nil {
 		log.Fatal(err)

src/config/config.go

@@ -10,8 +10,11 @@ import (
 
 const (
 	DefaultModelsDir            = "models"
+	DefaultVADThreshold         = 0.5
 	DefaultOverlayFontFamily    = "Sans"
 	MinimumOverlayFinalLifetime = 3 * time.Second
+	DefaultModelTimeout         = 30 * time.Second
+	DefaultModelShutdownTimeout = 2 * time.Second
 )
 
 type Mode string
@@ -27,6 +30,10 @@ type Settings struct {
 	ModelsDir            string
 	Threads              int
 	PreviewThresholdDB   float64
+	VADThreshold         float64
+	ModelAutoRestart     bool
+	ModelTimeout         time.Duration
+	ModelShutdownTimeout time.Duration
 }
 
 func Parse() Settings {
@@ -36,7 +43,11 @@ func Parse() Settings {
 	flag.DurationVar(&settings.ChunkDuration, "chunk-duration", time.Second, "fixed chunk size or VAD preview refresh interval")
 	flag.StringVar(&settings.ModelsDir, "models-dir", DefaultModelsDir, "directory containing downloaded Sherpa models")
 	flag.IntVar(&settings.Threads, "threads", 2, "CPU threads used by recognition and VAD")
-	flag.Float64Var(&settings.PreviewThresholdDB, "preview-threshold-dbfs", -45, "RMS dBFS threshold used to begin VAD previews")
+	flag.Float64Var(&settings.PreviewThresholdDB, "preview-threshold-dbfs", -45, "RMS dBFS gate below which recognition stays idle")
+	flag.Float64Var(&settings.VADThreshold, "vad-threshold", DefaultVADThreshold, "speech detection sensitivity from 0 to 1; higher rejects more non-speech")
+	flag.BoolVar(&settings.ModelAutoRestart, "model-auto-restart", true, "restart the isolated model worker after a crash or timeout")
+	flag.DurationVar(&settings.ModelTimeout, "model-timeout", DefaultModelTimeout, "maximum model startup or processing time; zero disables hang detection")
+	flag.DurationVar(&settings.ModelShutdownTimeout, "model-shutdown-timeout", DefaultModelShutdownTimeout, "time allowed for model flush and cleanup before force-killing the worker")
 	flag.Parse()
 	settings.Mode = Mode(mode)
 	return settings
@@ -86,7 +97,11 @@ func ParseDesktop() DesktopSettings {
 	flags.DurationVar(&desktop.ChunkDuration, "chunk-duration", time.Second, "fixed chunk size or VAD preview refresh interval")
 	flags.StringVar(&desktop.ModelsDir, "models-dir", DefaultModelsDir, "directory containing downloaded Sherpa models")
 	flags.IntVar(&desktop.Threads, "threads", 2, "CPU threads used by recognition and VAD")
-	flags.Float64Var(&desktop.PreviewThresholdDB, "preview-threshold-dbfs", -45, "RMS dBFS threshold used to begin VAD previews")
+	flags.Float64Var(&desktop.PreviewThresholdDB, "preview-threshold-dbfs", -45, "RMS dBFS gate below which recognition stays idle")
+	flags.Float64Var(&desktop.VADThreshold, "vad-threshold", DefaultVADThreshold, "speech detection sensitivity from 0 to 1; higher rejects more non-speech")
+	flags.BoolVar(&desktop.ModelAutoRestart, "model-auto-restart", true, "restart the isolated model worker after a crash or timeout")
+	flags.DurationVar(&desktop.ModelTimeout, "model-timeout", DefaultModelTimeout, "maximum model startup or processing time; zero disables hang detection")
+	flags.DurationVar(&desktop.ModelShutdownTimeout, "model-shutdown-timeout", DefaultModelShutdownTimeout, "time allowed for model flush and cleanup before force-killing the worker")
 	flags.StringVar(&output, "output", "", "caption output: terminal, overlay, or both (required)")
 	flags.StringVar(&backend, "overlay-backend", "auto", "display backend: auto, wayland, or x11")
 	flags.Float64Var(&desktop.Overlay.Opacity, "overlay-opacity", 0.90, "caption background opacity from 0 to 1")
@@ -114,6 +129,15 @@ func (s Settings) Validate() error {
 	if s.Threads <= 0 {
 		return errors.New("--threads must be positive")
 	}
+	if s.VADThreshold < 0 || s.VADThreshold > 1 {
+		return errors.New("--vad-threshold must be between 0 and 1")
+	}
+	if s.ModelTimeout < 0 {
+		return errors.New("--model-timeout cannot be negative")
+	}
+	if s.ModelShutdownTimeout <= 0 {
+		return errors.New("--model-shutdown-timeout must be positive")
+	}
 	return nil
 }
 

src/config/config_test.go

@@ -6,7 +6,12 @@ import (
 )
 
 func TestSettingsValidate(t *testing.T) {
-	valid := Settings{Mode: ModeFixed, ChunkDuration: time.Second, Threads: 1}
+	valid := Settings{
+		Mode:                 ModeFixed,
+		ChunkDuration:        time.Second,
+		Threads:              1,
+		ModelShutdownTimeout: DefaultModelShutdownTimeout,
+	}
 	if err := valid.Validate(); err != nil {
 		t.Fatalf("valid settings failed: %v", err)
 	}
@@ -17,7 +22,13 @@ func TestSettingsValidate(t *testing.T) {
 
 func validDesktopSettings() DesktopSettings {
 	return DesktopSettings{
-		Settings: Settings{Mode: ModeVAD, ChunkDuration: time.Second, Threads: 2},
+		Settings: Settings{
+			Mode:                 ModeVAD,
+			ChunkDuration:        time.Second,
+			Threads:              2,
+			ModelTimeout:         DefaultModelTimeout,
+			ModelShutdownTimeout: DefaultModelShutdownTimeout,
+		},
 		Output: OutputOverlay,
 		Overlay: OverlaySettings{
 			Backend: BackendAuto, Opacity: 0.9, FontSize: 28, FontFamily: DefaultOverlayFontFamily, BottomMargin: 100,
@@ -46,6 +57,10 @@ func TestDesktopSettingsValidate(t *testing.T) {
 		{"empty monitor", func(s *DesktopSettings) { s.Overlay.Monitor = "" }},
 		{"negative timeout", func(s *DesktopSettings) { s.Overlay.FinalTimeout = -time.Second }},
 		{"short timeout", func(s *DesktopSettings) { s.Overlay.FinalTimeout = time.Second }},
+		{"negative model timeout", func(s *DesktopSettings) { s.ModelTimeout = -time.Second }},
+		{"zero model shutdown timeout", func(s *DesktopSettings) { s.ModelShutdownTimeout = 0 }},
+		{"negative VAD threshold", func(s *DesktopSettings) { s.VADThreshold = -0.1 }},
+		{"large VAD threshold", func(s *DesktopSettings) { s.VADThreshold = 1.1 }},
 	}
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {

src/models/sherpa.go

@@ -25,6 +25,11 @@ type Resources struct {
 	vad        *sherpa.VoiceActivityDetector
 }
 
+func Validate(settings config.Settings) error {
+	_, err := validatePaths(settings.ModelsDir, settings.Mode)
+	return err
+}
+
 func New(settings config.Settings) (*Resources, error) {
 	paths, err := validatePaths(settings.ModelsDir, settings.Mode)
 	if err != nil {
@@ -52,7 +57,7 @@ func New(settings config.Settings) (*Resources, error) {
 
 	resources := &Resources{recognizer: recognizer}
 	if settings.Mode == config.ModeVAD {
-		resources.vad = newVAD(paths.vad, settings.Threads)
+		resources.vad = newVAD(paths.vad, settings.Threads, settings.VADThreshold)
 		if resources.vad == nil {
 			resources.Close()
 			return nil, fmt.Errorf("create Silero VAD: Sherpa returned nil")
@@ -87,6 +92,10 @@ func (r *Resources) AcceptVAD(samples []float32) {
 	r.vad.AcceptWaveform(samples)
 }
 
+func (r *Resources) SpeechActive() bool {
+	return r.vad.IsSpeech()
+}
+
 func (r *Resources) FlushVAD() {
 	r.vad.Flush()
 }
@@ -127,11 +136,11 @@ func validatePaths(modelsDir string, mode config.Mode) (modelPaths, error) {
 	return paths, nil
 }
 
-func newVAD(model string, threads int) *sherpa.VoiceActivityDetector {
+func newVAD(model string, threads int, threshold float64) *sherpa.VoiceActivityDetector {
 	return sherpa.NewVoiceActivityDetector(&sherpa.VadModelConfig{
 		SileroVad: sherpa.SileroVadModelConfig{
 			Model:     model,
-			Threshold: 0.5,
+			Threshold: float32(threshold),
 			// A 500 ms pause ends the current caption line and starts the next one.
 			MinSilenceDuration: 0.5,
 			MinSpeechDuration:  0.25,