SubtitleEdit/subtitleedit · error · InvalidOperationException

OmniVoice (CrispASR) — the crispasr server crashed during sy

Error message

OmniVoice (CrispASR) — the crispasr server crashed during synthesis.

What it means

Thrown inside the catch(HttpRequestException) handler in OmniVoiceCrispAsr.Speak when the POST to the local crispasr omnivoice server fails AND _serverProcess?.HasExited is true (the 'died' branch). This means the server process crashed or was killed during the synthesis request, and the connection drop is a consequence of the process death. StopServerInternal is called to clean up before throwing.

Source

Thrown at src/ui/Features/Video/TextToSpeech/Engines/OmniVoiceCrispAsr.cs:469

            response = await HttpClient.PostAsync($"{ServerBaseUrl}/v1/audio/speech", content, cancellationToken);
        }
        catch (HttpRequestException ex)
        {
            var serverLog = SnapshotServerLog();
            var launchCommand = _serverLaunchCommand;
            var died = _serverProcess?.HasExited == true;
            if (died)
            {
                StopServerInternal();
            }

            var failMsg = $"OmniVoice (CrispASR) request failed — Voice: {omniVoice}, Text: {text}, "
                + $"RequestJson: {body}, ServerExited: {died}, ServerLog: {serverLog}"
                + LaunchCmdSuffix(launchCommand);
            Se.LogError(ex, failMsg);
            Se.WriteToolsLog(failMsg);

            throw new InvalidOperationException(
                (died
                    ? "OmniVoice (CrispASR) — the crispasr server crashed during synthesis."
                    : "OmniVoice (CrispASR) request failed — the connection to the crispasr server was dropped.")
                + (string.IsNullOrEmpty(serverLog) ? string.Empty : $"{Environment.NewLine}Server log:{Environment.NewLine}{serverLog}")
                + LaunchCmdSuffix(launchCommand),
                ex);
        }

        using (response)
        {
            if (!response.IsSuccessStatusCode)
            {
                var errorBody = await SafeReadErrorAsync(response, cancellationToken);
                var serverLog = SnapshotServerLog();
                var launchCommand = _serverLaunchCommand;
                var errMsg = $"OmniVoice (CrispASR) server error {(int)response.StatusCode} {response.StatusCode} — "
                    + $"Voice: {omniVoice}, Text: {text}, RequestJson: {body}, "
                    + $"ResponseBody: {errorBody}, ServerLog: {serverLog}"

View on GitHub (pinned to 17a9f07487)

Solutions

Read the 'Server log' in the exception — native crashes often log a signal (SIGSEGV, SIGABRT) or an assertion before dying.
Re-import the reference voice WAV to ensure it is a valid 24 kHz mono file (ffmpeg conversion on import should handle this, but a manually-copied file may not).
Ensure only the F16 tokenizer (omnivoice-tokenizer-f16.gguf) is in the models folder — remove any q8_0 tokenizer to prevent the codec-poisoning crash documented in the class remarks.
Update GPU drivers and/or switch to a CPU build of crispasr to rule out a Metal/CUDA-specific crash.
Switch to the Q4_K model quant to reduce memory pressure if the crash is OOM-related.

Example fix

// before — the server crashes silently during a long batch
await engine.Speak(text, ...);

// after — detect the crash and respawn with a fallback quant
try { await engine.Speak(text, ...); }
catch (InvalidOperationException ex) when (ex.Message.Contains("crashed during synthesis"))
{
    OmniVoiceCrispAsr.StopServer();
    // next Speak() call will re-launch with fresh state
    throw new TtsRetryableException(ex);
}

Defensive patterns

Strategy: retry

Validate before calling

// Before synthesis, check server health and reference voice validity
if (_serverProcess is { HasExited: true })
    OmniVoiceCrispAsr.StopServer(); // clear stale state for respawn
if (!string.IsNullOrEmpty(voicePath) && !File.Exists(voicePath))
    throw new FileNotFoundException("Reference voice WAV missing: " + voicePath);

Type guard

null

Try / catch

catch (InvalidOperationException ex) when (ex.Message.Contains("crashed during synthesis"))
{
    // The server process died. OmniVoiceCrispAsr.StopServer() was already called internally.
    // Retry once — EnsureServerRunningAsync will launch a fresh server.
    if (attempt < maxRetries) { OmniVoiceCrispAsr.StopServer(); goto retry; }
    throw;
}

Prevention

Ensure only the F16 tokenizer is in the models folder (remove q8_0 to prevent codec crashes).
Re-import reference voice WAVs to guarantee valid 24 kHz mono format.
Monitor VRAM usage during batch synthesis to catch OOM before it crashes the server.

When it happens

Trigger: HttpClient.PostAsync raises HttpRequestException, and the separate check _serverProcess.HasExited returns true. This happens when the omnivoice backend segfaults during inference, is killed by the OS OOM-killer, or hits a native crash (e.g., Metal/CUDA driver fault) mid-synthesis.

Common situations: Synthesizing with a corrupted reference voice WAV that causes a native crash in the codec encoder; VRAM exhaustion on a GPU that was also running other workloads; the q8_0 tokenizer was accidentally placed in the models folder and crispasr picked it up despite the --codec-model pin (the class doc warns this produces garbage then can crash); a macOS Metal driver panic on certain model layers.

Related errors

AI-assisted analysis of SubtitleEdit/subtitleedit@17a9f07487 (2026-08-13). Data as JSON: /api/errors/403aa386c5f12fbf. Report an issue: GitHub.