refactor: move source files to src/ directory and rename project to Captioneer

Restructure Go source code into a dedicated `src/` directory, rename the project from "audio-listen" to "Captioneer", and update README accordingly. Add `.gitignore` entry for the built binary `captioneer`.
This commit is contained in:
2026-07-16 12:01:51 +03:00
parent d042637351
commit ce2161b7e6
12 changed files with 385 additions and 327 deletions
+51
View File
@@ -0,0 +1,51 @@
package main
import (
"encoding/binary"
"fmt"
"math"
"time"
)
const (
sampleRate = 16000
channels = 1
bytesPerSample = 2
packetDuration = 100 * time.Millisecond
)
func samplesPerPacket() int {
return durationSamples(packetDuration)
}
func pcm16LEToFloat32(pcm []byte) []float32 {
samples := make([]float32, len(pcm)/bytesPerSample)
for i := range samples {
samples[i] = float32(int16(binary.LittleEndian.Uint16(pcm[i*2:]))) / 32768
}
return samples
}
func durationSamples(duration time.Duration) int {
return int(duration * sampleRate / time.Second)
}
func rmsDBFS(samples []float32) float64 {
if len(samples) == 0 {
return math.Inf(-1)
}
var sum float64
for _, sample := range samples {
sum += float64(sample * sample)
}
rms := math.Sqrt(sum / float64(len(samples)))
if rms == 0 {
return math.Inf(-1)
}
return 20 * math.Log10(rms)
}
func formatAudioTime(samples int) string {
return fmt.Sprintf("%.2fs", float64(samples)/sampleRate)
}
+40
View File
@@ -0,0 +1,40 @@
package main
import (
"encoding/binary"
"math"
"testing"
"time"
)
func TestPCM16LEToFloat32(t *testing.T) {
pcm := make([]byte, 6)
binary.LittleEndian.PutUint16(pcm[0:], uint16(0))
binary.LittleEndian.PutUint16(pcm[2:], uint16(32767))
binary.LittleEndian.PutUint16(pcm[4:], uint16(32768))
got := pcm16LEToFloat32(pcm)
want := []float32{0, 32767.0 / 32768.0, -1}
for i := range want {
if math.Abs(float64(got[i]-want[i])) > 0.00001 {
t.Fatalf("sample %d = %f, want %f", i, got[i], want[i])
}
}
}
func TestDurationSamples(t *testing.T) {
if got := durationSamples(time.Second); got != sampleRate {
t.Fatalf("one second = %d samples, want %d", got, sampleRate)
}
if got := durationSamples(1500 * time.Millisecond); got != 24000 {
t.Fatalf("1.5 seconds = %d samples, want 24000", got)
}
}
func TestRMSDBFS(t *testing.T) {
if got := rmsDBFS([]float32{1, -1}); math.Abs(got) > 0.00001 {
t.Fatalf("full-scale RMS = %f dBFS, want 0", got)
}
if got := rmsDBFS([]float32{0, 0}); !math.IsInf(got, -1) {
t.Fatalf("silence = %f dBFS, want -Inf", got)
}
}
+129
View File
@@ -0,0 +1,129 @@
package main
import (
"fmt"
"io"
"strings"
sherpa "github.com/k2-fsa/sherpa-onnx-go-linux"
)
type captionProcessor struct {
cfg settings
recognizer *sherpa.OfflineRecognizer
vad *sherpa.VoiceActivityDetector
output io.Writer
fixedSamples []float32
previewSamples []float32
previewActive bool
previewVisible bool
nextPreviewAt int
totalSamples int
}
func newCaptionProcessor(cfg settings, recognizer *sherpa.OfflineRecognizer, vad *sherpa.VoiceActivityDetector, output io.Writer) *captionProcessor {
return &captionProcessor{
cfg: cfg,
recognizer: recognizer,
vad: vad,
output: output,
nextPreviewAt: durationSamples(cfg.chunkDuration),
}
}
func (p *captionProcessor) accept(samples []float32) {
if len(samples) == 0 {
return
}
if p.cfg.mode == "fixed" {
p.acceptFixed(samples)
return
}
p.acceptVAD(samples)
}
func (p *captionProcessor) acceptFixed(samples []float32) {
p.fixedSamples = append(p.fixedSamples, samples...)
chunkSize := durationSamples(p.cfg.chunkDuration)
for len(p.fixedSamples) >= chunkSize {
p.emitCommitted(p.totalSamples, p.totalSamples+chunkSize, p.fixedSamples[:chunkSize])
p.fixedSamples = p.fixedSamples[chunkSize:]
p.totalSamples += chunkSize
}
}
func (p *captionProcessor) acceptVAD(samples []float32) {
p.totalSamples += len(samples)
p.vad.AcceptWaveform(samples)
if p.previewActive || rmsDBFS(samples) >= p.cfg.previewThresholdDB {
p.previewActive = true
p.previewSamples = append(p.previewSamples, samples...)
if len(p.previewSamples) >= p.nextPreviewAt {
p.emitPreview(p.previewSamples)
p.nextPreviewAt += durationSamples(p.cfg.chunkDuration)
}
}
p.drainVAD()
}
func (p *captionProcessor) drainVAD() {
for !p.vad.IsEmpty() {
segment := p.vad.Front()
p.vad.Pop()
p.clearPreview()
p.emitCommitted(segment.Start, segment.Start+len(segment.Samples), segment.Samples)
p.previewSamples = nil
p.previewActive = false
p.nextPreviewAt = durationSamples(p.cfg.chunkDuration)
}
}
func (p *captionProcessor) flush() {
if p.cfg.mode == "fixed" {
if len(p.fixedSamples) > 0 {
p.emitCommitted(p.totalSamples, p.totalSamples+len(p.fixedSamples), p.fixedSamples)
}
return
}
p.vad.Flush()
p.drainVAD()
p.clearPreview()
}
func (p *captionProcessor) emitPreview(samples []float32) {
text := p.decode(samples)
if text == "" {
return
}
fmt.Fprintf(p.output, "\r\033[2K… %s", text)
p.previewVisible = true
}
func (p *captionProcessor) emitCommitted(start, end int, samples []float32) {
text := p.decode(samples)
if text == "" {
return
}
p.clearPreview()
fmt.Fprintf(p.output, "[%s-%s] %s\n", formatAudioTime(start), formatAudioTime(end), text)
}
func (p *captionProcessor) clearPreview() {
if p.previewVisible {
fmt.Fprint(p.output, "\r\033[2K")
p.previewVisible = false
}
}
func (p *captionProcessor) decode(samples []float32) string {
if len(samples) == 0 {
return ""
}
stream := sherpa.NewOfflineStream(p.recognizer)
defer sherpa.DeleteOfflineStream(stream)
stream.AcceptWaveform(sampleRate, samples)
p.recognizer.Decode(stream)
return strings.TrimSpace(stream.GetResult().Text)
}
+93
View File
@@ -0,0 +1,93 @@
package main
import (
"context"
"errors"
"fmt"
"io"
"os"
"os/exec"
"strings"
"sync"
)
func validatePrograms() error {
for _, program := range []string{"pactl", "parec"} {
if _, err := exec.LookPath(program); err != nil {
return fmt.Errorf("%s was not found in PATH: %w", program, err)
}
}
return nil
}
func defaultMonitorSource(ctx context.Context) (string, error) {
output, err := exec.CommandContext(ctx, "pactl", "get-default-sink").Output()
if err != nil {
return "", err
}
return strings.TrimSpace(string(output)) + ".monitor", nil
}
func capturePackets(ctx context.Context, monitorSource string) (<-chan []float32, func() error, error) {
cmd := exec.CommandContext(ctx, "parec",
"--device="+monitorSource,
"--format=s16le",
fmt.Sprintf("--rate=%d", sampleRate),
fmt.Sprintf("--channels=%d", channels),
"--raw",
"--latency-msec=50",
)
cmd.Stderr = os.Stderr
stdout, err := cmd.StdoutPipe()
if err != nil {
return nil, nil, err
}
if err := cmd.Start(); err != nil {
return nil, nil, err
}
packets := make(chan []float32, 50) // Five seconds at 100 ms per packet.
var waitOnce sync.Once
var waitErr error
wait := func() error {
waitOnce.Do(func() { waitErr = cmd.Wait() })
return waitErr
}
go func() {
defer close(packets)
packetBytes := make([]byte, samplesPerPacket()*bytesPerSample)
for {
n, readErr := io.ReadFull(stdout, packetBytes)
if n > 0 {
enqueueLatest(packets, pcm16LEToFloat32(packetBytes[:n]))
}
if readErr != nil {
if ctx.Err() == nil && !errors.Is(readErr, io.EOF) && !errors.Is(readErr, io.ErrUnexpectedEOF) {
fmt.Fprintf(os.Stderr, "audio capture read error: %v\n", readErr)
}
return
}
}
}()
return packets, wait, nil
}
func enqueueLatest(packets chan []float32, packet []float32) {
select {
case packets <- packet:
return
default:
}
// Keep the newest audio so captions recover quickly after overload.
select {
case <-packets:
default:
}
select {
case packets <- packet:
default:
}
fmt.Fprintln(os.Stderr, "warning: caption processing overloaded; dropped 100 ms of oldest audio")
}
+41
View File
@@ -0,0 +1,41 @@
package main
import (
"errors"
"flag"
"time"
)
const defaultModelsDir = "models"
type settings struct {
mode string
chunkDuration time.Duration
modelsDir string
threads int
previewThresholdDB float64
}
func parseFlags() settings {
var cfg settings
flag.StringVar(&cfg.mode, "mode", "", "caption mode: vad or fixed (required)")
flag.DurationVar(&cfg.chunkDuration, "chunk-duration", time.Second, "fixed chunk size or VAD preview refresh interval")
flag.StringVar(&cfg.modelsDir, "models-dir", defaultModelsDir, "directory containing downloaded Sherpa models")
flag.IntVar(&cfg.threads, "threads", 2, "CPU threads used by recognition and VAD")
flag.Float64Var(&cfg.previewThresholdDB, "preview-threshold-dbfs", -45, "RMS dBFS threshold used to begin VAD previews")
flag.Parse()
return cfg
}
func validateSettings(cfg settings) error {
if cfg.mode != "fixed" && cfg.mode != "vad" {
return errors.New("--mode is required and must be either fixed or vad")
}
if cfg.chunkDuration <= 0 {
return errors.New("--chunk-duration must be positive")
}
if cfg.threads <= 0 {
return errors.New("--threads must be positive")
}
return nil
}
+26
View File
@@ -0,0 +1,26 @@
package main
import (
"strings"
"testing"
"time"
)
func TestValidateSettings(t *testing.T) {
if err := validateSettings(settings{mode: "fixed", chunkDuration: time.Second, threads: 1}); err != nil {
t.Fatalf("valid settings failed: %v", err)
}
if err := validateSettings(settings{mode: "", chunkDuration: time.Second, threads: 1}); err == nil {
t.Fatal("missing mode unexpectedly passed")
}
}
func TestValidateModelsExplainsHowToInstallThem(t *testing.T) {
_, err := validateModels(t.TempDir(), "vad")
if err == nil {
t.Fatal("missing models unexpectedly passed validation")
}
if !strings.Contains(err.Error(), "./scripts/download-models.sh") {
t.Fatalf("missing-model error did not provide setup command: %v", err)
}
}
+59
View File
@@ -0,0 +1,59 @@
package main
import (
"context"
"fmt"
"log"
"os"
"os/signal"
"syscall"
sherpa "github.com/k2-fsa/sherpa-onnx-go-linux"
)
func main() {
cfg := parseFlags()
if err := validateSettings(cfg); err != nil {
log.Fatal(err)
}
if err := validatePrograms(); err != nil {
log.Fatal(err)
}
paths, err := validateModels(cfg.modelsDir, cfg.mode)
if err != nil {
log.Fatal(err)
}
ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
defer stop()
recognizer := newRecognizer(paths, cfg.threads)
defer sherpa.DeleteOfflineRecognizer(recognizer)
var vad *sherpa.VoiceActivityDetector
if cfg.mode == "vad" {
vad = newVAD(paths.vad, cfg.threads)
defer sherpa.DeleteVoiceActivityDetector(vad)
}
monitorSource, err := defaultMonitorSource(ctx)
if err != nil {
log.Fatalf("find default output monitor: %v", err)
}
fmt.Fprintf(os.Stderr, "Capturing from: %s\n", monitorSource)
fmt.Fprintln(os.Stderr, "Press Ctrl+C to stop.")
packets, waitCapture, err := capturePackets(ctx, monitorSource)
if err != nil {
log.Fatalf("start system-audio capture: %v", err)
}
processor := newCaptionProcessor(cfg, recognizer, vad, os.Stdout)
for packet := range packets {
processor.accept(packet)
}
processor.flush()
if err := waitCapture(); err != nil && ctx.Err() == nil {
log.Fatalf("audio capture stopped unexpectedly: %v", err)
}
}
+89
View File
@@ -0,0 +1,89 @@
package main
import (
"fmt"
"log"
"os"
"path/filepath"
"strings"
sherpa "github.com/k2-fsa/sherpa-onnx-go-linux"
)
type modelPaths struct {
encoder string
decoder string
joiner string
tokens string
vad string
}
func validateModels(modelsDir, mode string) (modelPaths, error) {
parakeetDir := filepath.Join(modelsDir, "parakeet-tdt-v2")
paths := modelPaths{
encoder: filepath.Join(parakeetDir, "encoder.int8.onnx"),
decoder: filepath.Join(parakeetDir, "decoder.int8.onnx"),
joiner: filepath.Join(parakeetDir, "joiner.int8.onnx"),
tokens: filepath.Join(parakeetDir, "tokens.txt"),
vad: filepath.Join(modelsDir, "silero_vad_v5.onnx"),
}
required := []string{paths.encoder, paths.decoder, paths.joiner, paths.tokens}
if mode == "vad" {
required = append(required, paths.vad)
}
var missing []string
for _, path := range required {
info, err := os.Stat(path)
if err != nil || info.IsDir() {
missing = append(missing, path)
}
}
if len(missing) > 0 {
return modelPaths{}, fmt.Errorf("missing required model files:\n %s\nRun ./scripts/download-models.sh", strings.Join(missing, "\n "))
}
return paths, nil
}
func newRecognizer(paths modelPaths, threads int) *sherpa.OfflineRecognizer {
recognizer := sherpa.NewOfflineRecognizer(&sherpa.OfflineRecognizerConfig{
FeatConfig: sherpa.FeatureConfig{SampleRate: sampleRate, FeatureDim: 80},
ModelConfig: sherpa.OfflineModelConfig{
Transducer: sherpa.OfflineTransducerModelConfig{
Encoder: paths.encoder,
Decoder: paths.decoder,
Joiner: paths.joiner,
},
Tokens: paths.tokens,
NumThreads: threads,
Provider: "cpu",
ModelType: "nemo_transducer",
},
DecodingMethod: "greedy_search",
})
if recognizer == nil {
log.Fatal("create Parakeet recognizer: Sherpa returned nil")
}
return recognizer
}
func newVAD(model string, threads int) *sherpa.VoiceActivityDetector {
vad := sherpa.NewVoiceActivityDetector(&sherpa.VadModelConfig{
SileroVad: sherpa.SileroVadModelConfig{
Model: model,
Threshold: 0.5,
MinSilenceDuration: 0.5,
MinSpeechDuration: 0.25,
WindowSize: 512,
MaxSpeechDuration: 30,
},
SampleRate: sampleRate,
NumThreads: threads,
Provider: "cpu",
}, 60)
if vad == nil {
log.Fatal("create Silero VAD: Sherpa returned nil")
}
return vad
}