Qwen3-TTS audio dropout - missing middle portion of generated speech

[touchaponk]

When generating TTS with Qwen3-TTS models, the output audio is

• missing portions in the middle of the speech. The beginning and end of the audio are present, but there's a noticeable gap/dropout in the middle section.
• sometimes the beginning part of the speech is missing

Steps to Reproduce

Expected Behavior

Complete, continuous audio output of the full sentence.

Actual Behavior

Audio has a dropout/gap in the middle portion. The beginning ("KFC is...") and end ("...third floor") are audible, but there's missing audio in between.

Environment
Macbook Pro M5
OS: macOS 26.2
Python: 3.12.12
mlx: 0.30.3
mlx-audio: 0.3.1
mlx-lm: 0.30.5
Model: mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit

audio_000.wav

Additional Notes

Issue persists even with added punctuation: "KFC, is located on the third floor."
Tried different phrasings, model version, quantize variations, but the dropout still occurs
Voice used: "Aiden"

[Blaizzy]

Could provide the full sentence and reproduce code you used?

[touchaponk]

Just by running the cli, here's the one I used:
python3 -m mlx_audio.tts.generate --model mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit --text "KFC is on the third floor" --voice Aiden

It happens for all model size and variation

[dxcore35]

The Qwen-3-TTS models are still not optimized and 100% functional.

[Blaizzy]

mlx_audio.tts.generate --model mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit --text "KFC is on the third floor" --voice Aiden

Thanks! I'm going to reproduce this and share the results.

[Blaizzy]

The Qwen-3-TTS models are still fucked up :/

Hey @dxcore35

This is not helpful at all and not the type of vocabulary that I appreciate in this repo.

If something is broken, here what you can do:

1. Open an issue detailing
2. Share a reproducible example
3. Find the issue
4. Bonus: Find solution and share it

[Blaizzy]

Just by running the cli, here's the one I used: python3 -m mlx_audio.tts.generate --model mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit --text "KFC is on the third floor" --voice Aiden

It happens for all model size and variation

Hey @touchaponk

I made a fresh install of mlx-audio v0.3.1 (latest) on a new UV env and run the same exact same model and command, and it worked fine. You can check the video below

Could you ensure you are running the correct env with the latest release, because the silence issue was addressed in v0.3.0 (PR #444).

Or if possible share another reproducible example.

customvoice.mp4

[kdu3142]

I'm having the same exact issue. Tried reinstalling everything from scratch, but no luck. Will try to reproduce it on my side and share an example.

[kdu3142]

@Blaizzy Here's my example. (Hope it helps, I'm sort of new at this)

Screen.Recording.2026-02-02.at.21.58.57.mp4

[touchaponk]

I can confirm that I'm running on 0.3.1

here's a new one from the cli:

audio_001.wav

[kdu3142]

I tested on my M4 and it seems to have worked. Might be M5 specific.

[touchaponk]

Some additional info that might be helpful is that there's a significant inference speed improvement for M5 when upgraded to macOS 26.2 : https://machinelearning.apple.com/research/exploring-llms-mlx-m5
So this could be one of the impact from apple's optimization (i.e. the neural accelerator) specific to M5 chip.
Unfortunately I've never run the tts generation before upgrading to macOS 26.2 so I'm not able to validate this.

[madeinoz67]

seeing similar issue on M1 ultra, finding it stutters or pauses between chunks. 64GB ram so is not a resource issue. Kokoro smooth as.

[Blaizzy]

seeing similar issue on M1 ultra, finding it stutters or pauses between chunks. 64GB ram so is not a resource issue. Kokoro smooth as.

Hey @madeinoz67

Thanks for reporting but unfortunately this is an unrelated issue.

That's completely normal for streaming.

You just need to adjust the streaming interval to your hardware or use a smaller quantization model (I.e., 4bit) or model (0.6B variant)

[alialamiidrissi]

I’m also running it on an M5 MacBook Pro. Force reinstalling from the main branch instead of the released package somehow fixed the issue.
uv pip install --reinstall git+https://github.com/Blaizzy/mlx-audio.git --prerelease=allow

[lucasnewman]

If it's M5 specific, I'm guessing the NAX fixes (e.g. ml-explore/mlx#3083) in MLX would resolve this, so it should be resolved with pip install -U mlx. If you're still seeing issues on M5 machines after that, please reopen.