pocket-tts

The manifest paths are usually fine. The length-bucketing loader does not emit batches until each loader shard contains at least num_bucket_batches × batch_size samples. With the old defaults, this required 20 × 64 × 6 = 7,680 utterances per GPU. Smaller shards produced the misleading 0 failures error.

Upgrade to a version containing PR #342, which lowers the defaults to num_bucket_batches: 5 and loader_procs: 3 and validates train/validation manifests at startup.

For very small finetuning datasets, disable bucketing and reduce loader processes in finetune.yaml:

data:
  num_bucket_batches: 1  # 1 disables length bucketing
  loader_procs: 1
  batch_size: 8          # must be no larger than the smallest train/validation shard

The requirement per GPU is:

num_bucket_batches × batch_size × loader_procs

With num_bucket_batches: 1 and loader_procs: 1, each split only needs at least batch_size readable entries. Make sure the validation manifest also meets this size because validation uses the training batch size; otherwise the run fails at the first validation step. Alternatively, add more validation samples or increase the dataset size.

This is caused by early EOS termination: for some French voices, the EOS logit can reach about -3 during the leading pause, crossing the default -4 threshold before speech is generated, so short texts end prematurely.

Fix by updating to the release containing the PR **"Ignore EOS on the first 6 generated frames"** and the new French models. Then use either:

- --language french — loads the new 6-layer French model
- french_24l — loads the new 24-layer French model

Immediate workaround if you cannot update: never synthesize fewer than three words. In streaming mode, concatenate a too-short chunk with the next chunk before synthesis. This reduced the reported failure rate from about 40% for 1–2 words to about 2% for 3+ words.

Adding a lead-in filler word did not help, and quantize=True made the issue worse on Apple Silicon.

Those shipped YAML configs omit default_temperature, so TTSModel.load_model() falls back to the schema default of 0.7 instead of the tuned English value 0.3. The fix changes the schema default to 0.3, making all shipped models sample at 0.3 while still allowing per-config overrides. After upgrading to a release containing the fix for #322, verify with:

m = TTSModel.load_model(language="spanish")
print(m.temp)  # should be 0.3

default_temperature: 0.3

Upgrade to a Pocket TTS release containing PR #228. The fix creates a per-request threading.Event, passes it into generate_audio_stream(), and sets it when the client disconnects. The generation loop in generate_data_with_state checks this event and exits through its normal termination path instead of decoding the remaining text. If applying the fix manually: 1) create stop_event = threading.Event() for each request, 2) pass it to generate_audio_stream(..., stop_event=stop_event), 3) set stop_event.set() in the disconnect/cancellation handler, 4) check the event inside the generation loop and stop before running another tts_model request. This prevents wasted CPU and avoids concurrent use of the non-thread-safe model; reported interrupt-to-next-reply latency drops from about 7.9 s to 0.27 s.

The failure can be caused by a stale HF_TOKEN environment variable. Remove it and run the export again.

PowerShell (current terminal):

Remove-Item Env:HF_TOKEN

To remove it permanently from your Windows user environment:

[Environment]::SetEnvironmentVariable("HF_TOKEN", $null, "User")

Restart your terminal, verify it is empty with echo $env:HF_TOKEN, then rerun:

pocket-tts export-voice .\turtle_sample.wav turtle.safetensors

If voice cloning requires authentication, replace the stale token with a valid one from Hugging Face instead:

$env:HF_TOKEN = "hf_your_valid_token"