Kokoro not working

[kuanjames]

I am getting this error:

External: 500, message='Internal Server Error', url='http://192.168.0.10:8000/v1/audio/speech'

[kuanjames]

Log:

urllib3.connectionpool - DEBUG - [-] - https://raw.githubusercontent.com:443 "GET /explosion/spacy-models/master/compatibility.json HTTP/1.1" 200 4359
2026-04-03 21:03:31,242 - omlx.engine.tts - INFO - [-] - TTS synthesize: model=/Users/xxx/.lmstudio/models/Kokoro-82M-bf16, text_len=103, voice=af_heart, speed=1.0
2026-04-03 21:03:31,541 - urllib3.connectionpool - DEBUG - [-] - Starting new HTTPS connection (1): raw.githubusercontent.com:443
2026-04-03 21:03:31,679 - urllib3.connectionpool - DEBUG - [-] - https://raw.githubusercontent.com:443 "GET /explosion/spacy-models/master/compatibility.json HTTP/1.1" 200 4359

[shivarajd]

Terminal:

curl http://127.0.0.1:8003/v1/audio/speech
-H "Content-Type: application/json"
-d '{"model": "Kokoro-82M-bf16", "speed": "0.8", "voice": "am_eric", "input": "Hello world, I am Eric"}'
--output speech.wav && afplay speech.wav
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 120 0 21 100 99 26 123 --:--:-- --:--:-- --:--:-- 149
Error: AudioFileOpen failed ('typ?')

Log:

2026-04-03 22:56:24,595 - omlx.engine_pool - INFO - [-] - Loading model: Kokoro-82M-bf16
2026-04-03 22:56:24,605 - omlx.engine.tts - INFO - [-] - Starting TTS engine: /Users/shiva/.omlx/models/Kokoro-82M-bf16
2026-04-03 22:56:39,798 - omlx.engine.tts - INFO - [-] - TTS engine started: /Users/shiva/.omlx/models/Kokoro-82M-bf16
2026-04-03 22:56:39,867 - omlx.engine_pool - INFO - [-] - Loaded model: Kokoro-82M-bf16 (estimated: 327.56MB, total: 3.42GB)
2026-04-03 22:56:39,869 - omlx.engine.tts - INFO - [-] - TTS synthesize: model=/Users/shiva/.omlx/models/Kokoro-82M-bf16, text_len=22, voice=am_eric, speed=0.8

Other TTS models, chatterbox-turbo-4bit, kitten-tts-mini-0.8-bf16, and pocket-tts-8bit works well.

[jundot]

Fixed in 1464be6.

The issue was that Kokoro-82M-bf16 returns audio data in bfloat16 dtype, which numpy doesn't support natively. The np.array() conversion was failing silently and getting caught as a 500 error.

The fix casts bfloat16 to float32 before the numpy conversion in audio_to_wav_bytes(). This is a shared utility so it covers TTS, STT, and STS engines.

Please update to the latest version and let me know if the issue persists.

[shivarajd]

The issue still persists with Kokoro-82M-bf16 in oMLX v0.3.3.

curl http://127.0.0.1:8003/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "Kokoro-82M-bf16", "voice": "am_eric", "speed": "0.8", "input": "Hello world, I am Eric"}' \
  --output speech.wav && afplay speech.wav
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100   120    0    21  100    99     19     92  0:00:01  0:00:01 --:--:--   111
Error: AudioFileOpen failed ('typ?')

Log:

2026-04-05 07:55:13,782 - omlx.engine_pool - INFO - [-] - Loading model: Kokoro-82M-bf16
2026-04-05 07:55:13,783 - omlx.engine.tts - INFO - [-] - Starting TTS engine: /Users/shiva/.omlx/models/Kokoro-82M-bf16
2026-04-05 07:55:14,704 - omlx.engine.tts - INFO - [-] - TTS engine started: /Users/shiva/.omlx/models/Kokoro-82M-bf16
2026-04-05 07:55:14,705 - omlx.engine_pool - INFO - [-] - Loaded model: Kokoro-82M-bf16 (estimated: 327.56MB, total: 327.56MB)
2026-04-05 07:55:14,705 - omlx.engine.tts - INFO - [-] - TTS synthesize: model=/Users/shiva/.omlx/models/Kokoro-82M-bf16, text_len=22, voice=am_eric, speed=0.8

Tested with kitten-tts-mini-0.8-bf16, pocket-tts-8bit, and chatterbox-turbo-4bit they all work fine.

kitten-tts-mini-0.8-bf16:

curl http://127.0.0.1:8003/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "kitten-tts-mini-0.8-bf16", "voice": "Leo", "input": "Hello world, I am Leo"}' \
  --output speech.wav && afplay speech.wav
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  140k  100  140k  100    87  98485     59  0:00:01  0:00:01 --:--:-- 98584

Log:

2026-04-05 07:57:36,586 - omlx.engine_pool - INFO - [-] - Loading model: kitten-tts-mini-0.8-bf16
2026-04-05 07:57:36,587 - omlx.engine.tts - INFO - [-] - Starting TTS engine: /Users/shiva/.omlx/models/kitten-tts-mini-0.8-bf16
2026-04-05 07:57:36,778 - omlx.engine.tts - INFO - [-] - TTS engine started: /Users/shiva/.omlx/models/kitten-tts-mini-0.8-bf16
2026-04-05 07:57:36,805 - omlx.engine_pool - INFO - [-] - Loaded model: kitten-tts-mini-0.8-bf16 (estimated: 289.50MB, total: 757.48MB)
2026-04-05 07:57:36,805 - omlx.engine.tts - INFO - [-] - TTS synthesize: model=/Users/shiva/.omlx/models/kitten-tts-mini-0.8-bf16, text_len=21, voice=Leo, speed=1.0
2026-04-05 07:57:38,037 - omlx.engine.tts - INFO - [-] - TTS synthesize done: model=/Users/shiva/.omlx/models/kitten-tts-mini-0.8-bf16, 1.23s, 144044 bytes output

pocket-tts-8bit:

curl http://127.0.0.1:8003/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "pocket-tts-8bit", "voice": "alba", "input": "Hello world, I am Alba"}' \
  --output speech.wav && afplay speech.wav
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  108k  100  108k  100    80   236k    173 --:--:-- --:--:-- --:--:--  237k

Log:

2026-04-05 07:57:26,351 - omlx.engine_pool - INFO - [-] - Loading model: pocket-tts-8bit
2026-04-05 07:57:26,352 - omlx.engine.tts - INFO - [-] - Starting TTS engine: /Users/shiva/.omlx/models/pocket-tts-8bit
2026-04-05 07:57:26,397 - mlx_audio.tts.models.pocket_tts.conditioners - INFO - [-] - Loading sentencepiece tokenizer from hf://kyutai/pocket-tts-without-voice-cloning/tokenizer.model@d4fdd22ae8c8e1cb3634e150ebeff1dab2d16df3
2026-04-05 07:57:26,482 - omlx.engine.tts - INFO - [-] - TTS engine started: /Users/shiva/.omlx/models/pocket-tts-8bit
2026-04-05 07:57:26,483 - omlx.engine_pool - INFO - [-] - Loaded model: pocket-tts-8bit (estimated: 140.42MB, total: 467.98MB)
2026-04-05 07:57:26,483 - omlx.engine.tts - INFO - [-] - TTS synthesize: model=/Users/shiva/.omlx/models/pocket-tts-8bit, text_len=22, voice=alba, speed=1.0
2026-04-05 07:57:26,807 - omlx.engine.tts - INFO - [-] - TTS synthesize done: model=/Users/shiva/.omlx/models/pocket-tts-8bit, 0.32s, 111404 bytes output

chatterbox-turbo-4bit:

curl http://127.0.0.1:8003/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d $'{"model": "chatterbox-turbo-4bit", "input": "[sigh] I can\'t believe it\'s Monday again. [groan] But hey, [clear throat] let\'s make the best of it!"}' \
  --output speech.wav && afplay speech.wav
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100  294k  100  294k  100   147  49968     24  0:00:06  0:00:06 --:--:-- 74966

Log:

2026-04-05 07:57:54,260 - omlx.engine_pool - INFO - [-] - Loading model: chatterbox-turbo-4bit
2026-04-05 07:57:54,260 - omlx.engine.tts - INFO - [-] - Starting TTS engine: /Users/shiva/.omlx/models/chatterbox-turbo-4bit
2026-04-05 07:57:55,827 - mlx_audio.tts.models.chatterbox_turbo.chatterbox_turbo - INFO - [-] - Loading 15 VE weights
2026-04-05 07:57:55,827 - mlx_audio.tts.models.chatterbox_turbo.chatterbox_turbo - INFO - [-] - Loading 508 T3 weights
2026-04-05 07:57:55,829 - mlx_audio.tts.models.chatterbox_turbo.chatterbox_turbo - INFO - [-] - Loading 2230 S3Gen weights
2026-04-05 07:57:56,242 - mlx_audio.tts.models.chatterbox_turbo.chatterbox_turbo - INFO - [-] - Weights loaded successfully
2026-04-05 07:57:56,296 - mlx_audio.tts.models.chatterbox_turbo.chatterbox_turbo - INFO - [-] - Loaded text tokenizer
2026-04-05 07:57:56,803 - httpx - INFO - [-] - HTTP Request: HEAD https://huggingface.co/mlx-community/S3TokenizerV2/resolve/main/model.safetensors "HTTP/1.1 302 Found"
2026-04-05 07:57:56,805 - huggingface_hub.utils._http - WARNING - [-] - Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
2026-04-05 07:57:56,816 - mlx_audio.tts.models.chatterbox_turbo.chatterbox_turbo - INFO - [-] - Loaded S3 speech tokenizer weights
2026-04-05 07:57:56,817 - mlx_audio.tts.models.chatterbox_turbo.chatterbox_turbo - INFO - [-] - Loaded pre-computed conditionals from safetensors
2026-04-05 07:57:56,819 - omlx.engine.tts - INFO - [-] - TTS engine started: /Users/shiva/.omlx/models/chatterbox-turbo-4bit
2026-04-05 07:57:56,845 - omlx.engine_pool - INFO - [-] - Loaded model: chatterbox-turbo-4bit (estimated: 415.72MB, total: 1.15GB)
2026-04-05 07:57:56,845 - omlx.engine.tts - INFO - [-] - TTS synthesize: model=/Users/shiva/.omlx/models/chatterbox-turbo-4bit, text_len=100, voice=None, speed=1.0
2026-04-05 07:58:00,270 - omlx.engine.tts - INFO - [-] - TTS synthesize done: model=/Users/shiva/.omlx/models/chatterbox-turbo-4bit, 3.42s, 301364 bytes output

[kuanjames]

@jundot this is still a problem in 0.3.4 and also it is a problem with 8bit too, not just bf16