python export_onnx.py no longer works

[mp20]

python export_onnx.py

PocketTTS ONNX Export

Downloading weights...
100% (224/224 MB)
✓ models/.cache/tts_b6369a24.safetensors
Downloading tokenizer...
100% (0/0 MB)
✓ models/tokenizer.model
Traceback (most recent call last):
File "/home/tiam/Downloads/PocketTTS.cpp/export_onnx.py", line 1317, in
main()
File "/home/tiam/Downloads/PocketTTS.cpp/export_onnx.py", line 1255, in main
config = load_config(config_path)
File "/home/tiam/Downloads/my_home_agent/.venv-pocket/lib/python3.10/site-packages/pocket_tts/utils/config.py", line 126, in load_config
raise FileNotFoundError(
FileNotFoundError: Config file not found: /Downloads/my_home_agent/.venv-pocket/lib/python3.10/site-packages/pocket_tts/config/b6369a24.yaml. Did you make a typo? Available languages: ['english_2026-01', 'italian', 'portuguese', 'spanish', 'french_24l', 'english_2026-04', 'english', 'portuguese_24l', 'spanish_24l', 'italian_24l', 'german_24l', 'german']

[rhodey]

My PR #7 adds docker support and also fixes this issue

[handpickencounter]

pocket-tts added a BOS token to the new models, not sure if it matters for export but it does matter for inference.

[rhodey]

@handpickencounter if you have any more info to add about this BOS token please share here. It sounds like you know some good info about this and I am planning to use PocketTTS.cpp soon in a production so if an issue shows up with this I will be committed to trying to fix it at which time I would be very grateful to have this added context info from you.

When I first opened PR #7 there were no noticeable issues with PocketTTS.cpp but then they started to add the non-english language support and now for the last few days (and after docker rebuild) when using PocketTTS.cpp #7 I sometimes get buggy voice which may be what #10 is describing but it may be related to this BOS token you explain.

The task I'm on right now is not to track down and fix the new PocketTTS.cpp sometimes buggy voice but this will be my task soon so again thank you if you can add anymore context about BOS

[handpickencounter]

I'm only using the export script from this repo as the result much more performant than the one by Kevin. my usecase is in python.

I used your patch but changed to 2026-04 English, it failed because it downloaded the wrong safetensors but i just downloaded the right one manually and replaced the one in .cache with it (renaming the downloaded model.safetensors to the one the script downloaded) as I was too lazy to fix the export script.

After doing that I was just missing the BOS token change (INFERENCE) which I fixed:

bos_before_voice = np.load("bos_before_voice.npy").astype(np.float32)  # https://huggingface.co/KevinAHM/pocket-tts-onnx/blob/main/onnx/english_2026-04/bos_before_voice.npy

embeddings = onnx_mimi_encoder.run(None, {"audio": audio})[0]
embeddings = np.concatenate([bos_before_voice, embeddings], axis=1)

that was all the change I needed to get the new english model working.