# What changed in mlx-audio

> 47 changes in mlx-audio 0.6 from its official release notes, upgrade guides and advisories: 14 new APIs and options. Each one cited. Sources checked 2026-10-28. HTML: https://unstale.sh/pypi/mlx-audio/

To see which of these apply to a repository, run `npx @unstale/cli init` in it.

## mlx-audio 0.6

### New APIs and options

- Adds ASR/AST model support. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Loads voices from the selected model snapshot. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Adds Ming 3.0 TTS, ASR, and Audio Tiny model support. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Adds STT model support. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Adds a speaker embedding extractor. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- A Gradium Phonon TTS model is added. (0.6.1) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.1)
- Support is added for the Ming-3.0-tts-0.5B model. (0.6.1) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.1)
- Adds support for Fish Audio Drama 3. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Adds the --expert-bits-map conversion option for fish_drama3. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Adds the --expert-group-size conversion option for fish_drama3. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Adds the --head-bits conversion option for fish_drama3. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Adds Granite Speech 4.1 Plus support. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Adds MiMo-V2.5-ASR speech-to-text support. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Adds Phonon-2 speech recognition support. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)

### Performance

- Bounds memory use across calls and for long chunks. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Breeze-TTS generation is optimized. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- VibeVoice receives performance improvements. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)

### Behavior and default changes

- encodec and phonon fall back to plain MLX when Metal is unavailable. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Model sampling parameters are now used by default. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)

### Fixes and other changes

- Flattens line breaks to avoid early turn ends. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Drops the <|semantic:0|> frame immediately before EOS. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Samples the main token without a top-k cap. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Fixes model loading and threaded generation. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Applies repetition_penalty to log-probabilities. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Loads the October 2026 checkpoint. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Preserves query dtype in sliding attention masks. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Honors stream=True by yielding one result per text fragment. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Fixes generation sampling. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Sets streaming chunk flags correctly in five models. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Falls back to the Qwen2.5 tokenizer when the local tokenizer is empty. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- Resolves the no_speech token for pre-large-v3 vocabularies. (0.6.2) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.2)
- The Ming dense prompt contract is fixed. (0.6.1) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.1)
- Fixes rollback crashes with independent CFG and dual conditioning in irodori-tts. (0.6.1) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.1)
- Server-side STT uploads use a writable temporary file. (0.6.1) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.1)
- Audio input now decodes bytes and mislabeled files and adds a duration limit. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Stopping no longer drops audio that has not reached the auto-start threshold. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Prevents fast_ paths and embeddings from being quantized, fixing the convert script. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Empty Qwen3-ASR audio is validated. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Language auto-detection is kept per chunk. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Streaming no longer fabricates timestamps. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Makes the 0.6B instruct guard reachable. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- The server uses the Qwen3-TTS generation config when sampling parameters are omitted. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Prefills full text to avoid slow speech. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Streaming detokenization is now tokenizer-safe. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Upsampling now uses the existing FIR resampler. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- The server now forwards the prompt to STT models on this endpoint. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)
- Empty non-speech token encodings are ignored. (0.6.0) — [release notes](https://github.com/Blaizzy/mlx-audio/releases/tag/v0.6.0)