mudler/LocalAI · error · Error

Web Audio API not available

Error message

Web Audio API not available

What it means

This error is printed by the cgo bridge in LocalAI's whisper backend when whisper_vad_detect_speech() (whisper.cpp Silero VAD) returns false for the submitted PCM buffer. whisper.cpp returns false from this call when the audio cannot be processed for speech detection — most commonly because the buffer is shorter than the minimum the VAD model accepts, because the global vctx was never initialized (load_model_vad not called or previously failed), or because the compute pass itself errored. It is a C-level integer return code surfaced through cgo, not a Go error value.

Source

Thrown at core/http/react-ui/src/hooks/useAudioPeaks.js:40

  const [duration, setDuration] = useState(0)
  const [error, setError] = useState(null)
  const [loading, setLoading] = useState(false)

  useEffect(() => {
    setPeaks(null)
    setDuration(0)
    setError(null)
    setLoading(false)
    if (!src) return
    let cancelled = false
    setLoading(true)

    async function decode() {
      try {
        const response = await fetch(src)
        const buf = await response.arrayBuffer()
        const ctx = getSharedAudioContext()
        if (!ctx) throw new Error('Web Audio API not available')
        const audioBuf = await ctx.decodeAudioData(buf.slice(0))
        if (cancelled) return
        const data = audioBuf.getChannelData(0)
        const step = Math.max(1, Math.floor(data.length / buckets))
        const result = new Float32Array(buckets)
        for (let i = 0; i < buckets; i++) {
          let peak = 0
          const start = i * step
          const end = Math.min(start + step, data.length)
          for (let j = start; j < end; j++) {
            const v = Math.abs(data[j])
            if (v > peak) peak = v
          }
          result[i] = peak
        }
        setPeaks(result)
        setDuration(audioBuf.duration)
        setLoading(false)

View on GitHub (pinned to 44413a9d06)

Solutions

  1. Check the audio length before calling vad(): whisper.cpp's Silero VAD requires a minimum buffer (>= 512 samples ≈ 32 ms per frame, and practically >= ~1 s for meaningful detection); buffer or concatenate short streaming chunks until you have at least ~0.5–1 s of 16 kHz mono PCM.
  2. Verify that load_model_vad() was called and returned 0 before vad(); inspect stderr for 'error: Failed to init model as VAD' — if present, the model path points at a non-VAD GGUF and must point at a Silero VAD model (e.g. silence-pc or silero VAD gguf).
  3. Confirm the PCM input is 16 kHz, mono, float32 normalized [-1,1]; re-run the pipeline with a known-good WAV file to isolate resampling bugs.
  4. Check the [ERROR]-prefixed ggml log lines (installed via whisper_log_set in the same file) immediately above this message — they name the underlying compute failure (backend load failure, OOM, unsupported tensor op); rebuild/re-run with CPU-only backends if a GPU backend is implicated.
  5. If whisper.cpp was recently updated, rebuild the gowhisper C++ shim against the matching headers so the whisper_vad_* ABI agrees with the vendored library.

Example fix

// Go caller: buffer short chunks before calling VAD
// before
segs, err := gw.VAD(chunk) // chunk may be 200 ms -> "error: failed to detect speech"

// after
const minVADSamples = 16000 // 1 s @ 16 kHz
buf = append(buf, chunk...)
if len(buf) < minVADSamples {
    continue // keep accumulating audio
}
segs, err := gw.VAD(buf)
if err != nil {
    log.Printf("vad failed (len=%d samples): %v", len(buf), err)
    return err
}
Defensive patterns

Strategy: validation

Validate before calling

// Run before calling vad(): minimum length, sample-rate, and init checks
const whisperSampleRate = 16000
const minVADSeconds = 1.0

func canRunVAD(sampleCount int, vadModelLoaded bool) error {
    if !vadModelLoaded { // load_model_vad() must have returned 0
        return errors.New("VAD model not loaded")
    }
    if sampleCount < int(minVADSeconds*whisperSampleRate) {
        return fmt.Errorf("audio too short for VAD: %d samples (< %d)",
            sampleCount, int(minVADSeconds*whisperSampleRate))
    }
    return nil
}

Type guard

// Go: shape check on the PCM slice handed to cgo
func isVADReadyPCM(pcm []float32) bool {
    return len(pcm) >= 16000 && // >= 1 s of 16 kHz mono
        !hasNaN(pcm)
}

func hasNaN(pcm []float32) bool {
    for _, s := range pcm {
        if math.IsNaN(s) || math.Abs(s) > 1e3 { // not normalized float PCM
            return true
        }
    }
    return false
}

Try / catch

// cgo has no exceptions: treat the integer return as the error
rc := C.vad((*C.float)(&pcm[0]), C.size_t(len(pcm)), &out, &outLen)
if rc == 1 {
    // do NOT retry the same buffer; log length and model state, then
    // drop or grow the buffer before the next attempt
    log.Printf("vad failed: len=%d samples, vctxOK=%v", len(pcm), vctxOK)
}

Prevention

When it happens

Trigger: Calling vad() with pcmf32_len below the Silero VAD minimum window (whisper.cpp rejects audio shorter than the model's minimum sample count, on the order of a few hundred ms at 16 kHz); calling vad() before a successful load_model_vad(), leaving the static vctx null (whisper.cpp then fails immediately); passing audio resampled to a rate other than 16 kHz mono float32; a ggml backend compute failure (e.g. incompatible GPU backend, out-of-memory) during the VAD forward pass; loading a regular transcription GGUF instead of a Silero-VAD GGUF would fail earlier at whisper_vad_init_from_file_with_params, but a truncated/corrupt VAD model can still init and then fail here.

Common situations: Users enabling LocalAI's VAD-based audio chunking and feeding very short clips (e.g. streaming chunks of 100–300 ms); forgetting to configure the VAD model path so the transcriber model is loaded but vctx is null; audio arriving stereo or 8 kHz because the upstream resampling step was skipped; whisper.cpp version drift between the vendored ggml/whisper sources and the prebuilt gowhisper library; GPU/CUDA or Metal backend failures on constrained hardware.

Related errors

AI-assisted analysis of mudler/LocalAI@44413a9d06 (2026-08-15). Data as JSON: /api/errors/ee67a33f05d057c7. Report an issue: GitHub.