pypi / mlx-audio
mlx-audio
What changed in mlx-audio 0.6, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-28.
14 new APIs and options3 performance2 behavior and defaults28 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initmlx-audio 0.647 changes
New APIs and options 14
Adds ASR/AST model support.
Loads voices from the selected model snapshot.
Adds Ming 3.0 TTS, ASR, and Audio Tiny model support.
Adds STT model support.
Adds a speaker embedding extractor.
A Gradium Phonon TTS model is added.
Support is added for the Ming-3.0-tts-0.5B model.
Adds support for Fish Audio Drama 3.
Adds the --expert-bits-map conversion option for fish_drama3.
Adds the --expert-group-size conversion option for fish_drama3.
Adds the --head-bits conversion option for fish_drama3.
Adds Granite Speech 4.1 Plus support.
Adds MiMo-V2.5-ASR speech-to-text support.
Adds Phonon-2 speech recognition support.
Performance 3
Bounds memory use across calls and for long chunks.
Breeze-TTS generation is optimized.
VibeVoice receives performance improvements.
Behavior and default changes 2
encodec and phonon fall back to plain MLX when Metal is unavailable.
Model sampling parameters are now used by default.
Fixes and other changes 28
Flattens line breaks to avoid early turn ends.
Drops the <|semantic:0|> frame immediately before EOS.
Samples the main token without a top-k cap.
Fixes model loading and threaded generation.
Applies repetition_penalty to log-probabilities.
Loads the October 2026 checkpoint.
Preserves query dtype in sliding attention masks.
Honors stream=True by yielding one result per text fragment.
Fixes generation sampling.
Sets streaming chunk flags correctly in five models.
Falls back to the Qwen2.5 tokenizer when the local tokenizer is empty.
Resolves the no_speech token for pre-large-v3 vocabularies.
The Ming dense prompt contract is fixed.
Fixes rollback crashes with independent CFG and dual conditioning in irodori-tts.
Server-side STT uploads use a writable temporary file.
Audio input now decodes bytes and mislabeled files and adds a duration limit.
Stopping no longer drops audio that has not reached the auto-start threshold.
Prevents fast_ paths and embeddings from being quantized, fixing the convert script.
Empty Qwen3-ASR audio is validated.
Language auto-detection is kept per chunk.
Streaming no longer fabricates timestamps.
Makes the 0.6B instruct guard reachable.
The server uses the Qwen3-TTS generation config when sampling parameters are omitted.
Prefills full text to avoid slow speech.
Streaming detokenization is now tokenizer-safe.
Upsampling now uses the existing FIR resampler.
The server now forwards the prompt to STT models on this endpoint.
Empty non-speech token encodings are ignored.