git.giacomolli.uk - onechat/blob - voice_transcriber.go

   1 package main
   2
   3 import (
   4         "context"
   5         "errors"
   6         "io"
   7         "os"
   8         "os/exec"
   9         "path/filepath"
  10         "strings"
  11         "time"
  12         "unicode/utf8"
  13 )
  14
  15 const (
  16         // voiceTranscriptionTimeout bounds one local transcription process.
  17         voiceTranscriptionTimeout = 2 * time.Minute
  18         voiceProcessExitTimeout   = time.Second
  19         voiceTranscriptMaxBytes   = 1 << 20
  20 )
  21
  22 // Nix sets these paths with linker flags. Other builds use config or PATH.
  23 var (
  24         voiceTranscriberPath  string
  25         voiceModelPath        string
  26         errVoiceUnavailable   = errors.New("transcriber or model is not available")
  27         errVoiceTranscription = errors.New("local transcription failed")
  28 )
  29
  30 // VoiceConfig selects a local executable and model. Explicit settings take
  31 // precedence over Nix paths and do not silently fall back when invalid.
  32 type VoiceConfig struct {
  33         Transcriber string `json:"transcriber"`
  34         Model       string `json:"model"`
  35 }
  36
  37 // VoiceTranscriber runs a local Whisper process. Process output and errors can
  38 // contain speech and private paths, so only fixed failure causes leave it.
  39 type VoiceTranscriber struct {
  40         config VoiceConfig
  41 }
  42
  43 // Text stores audio before decoding it. A failed decode keeps its voice marker,
  44 // so a missing model or bad audio does not stop the next message. A temporary
  45 // transcript stays in the private voice directory and is removed after use.
  46 func (v *VoiceTranscriber) Text(ctx context.Context, store *VoiceStore, data []byte) string {
  47         name, err := store.Save(data)
  48         if err != nil {
  49                 return "(voice not stored)"
  50         }
  51
  52         text, err := v.transcribe(ctx, filepath.Join(store.dir, name))
  53         if err != nil {
  54                 text = "(voice not transcribed: " + err.Error() + ")"
  55         }
  56
  57         return "[voice " + name + "]\n" + text
  58 }
  59
  60 func (v *VoiceTranscriber) transcribe(ctx context.Context, path string) (text string, err error) {
  61         program, model, err := voiceLookup(v.config, voiceTranscriberPath, voiceModelPath)
  62         if err != nil {
  63                 return "", errVoiceUnavailable
  64         }
  65
  66         ctx, cancel := context.WithTimeout(ctx, voiceTranscriptionTimeout)
  67         defer cancel()
  68
  69         dir, err := os.MkdirTemp(filepath.Dir(path), ".transcript-*")
  70         if err != nil {
  71                 return "", errVoiceTranscription
  72         }
  73
  74         defer func() {
  75                 if os.RemoveAll(dir) != nil {
  76                         err = errors.Join(err, errVoiceTranscription)
  77                 }
  78         }()
  79
  80         output := filepath.Join(dir, "text")
  81         command := exec.CommandContext(ctx, program, "--model", model, "--file", path,
  82                 "--language", "auto", "--max-context", "0", "--output-txt", "--output-file", output)
  83         command.Stdout = io.Discard
  84         command.Stderr = io.Discard
  85         command.WaitDelay = voiceProcessExitTimeout
  86         err = command.Run()
  87         if ctx.Err() != nil {
  88                 return "", ctx.Err()
  89         }
  90
  91         if err != nil {
  92                 return "", errVoiceTranscription
  93         }
  94
  95         file, err := os.Open(output + ".txt")
  96         if err != nil {
  97                 return "", errVoiceTranscription
  98         }
  99
 100         data, err := io.ReadAll(io.LimitReader(file, voiceTranscriptMaxBytes+1))
 101         err = errors.Join(err, file.Close())
 102         if err != nil || len(data) > voiceTranscriptMaxBytes || !utf8.Valid(data) {
 103                 return "", errVoiceTranscription
 104         }
 105
 106         text = strings.TrimSpace(string(data))
 107         if text == "" {
 108                 return "", errVoiceTranscription
 109         }
 110
 111         return text, nil
 112 }
 113
 114 func voiceLookup(config VoiceConfig, builtProgram, builtModel string) (string, string, error) {
 115         program := config.Transcriber
 116         if program == "" {
 117                 program = builtProgram
 118         }
 119
 120         if program == "" {
 121                 program = "whisper-cli"
 122         }
 123
 124         program, err := exec.LookPath(program)
 125         if err != nil {
 126                 return "", "", errVoiceUnavailable
 127         }
 128
 129         model := config.Model
 130         if model == "" {
 131                 model = builtModel
 132         }
 133
 134         info, err := os.Stat(model)
 135         if err != nil || !info.Mode().IsRegular() || info.Size() == 0 {
 136                 return "", "", errVoiceUnavailable
 137         }
 138
 139         return program, model, nil
 140 }