v0.4.5

What's Changed

• Add ZONOS2 TTS model by @lucasnewman in #791
• Cache Higgs v3 reference audio codes by @sxy-trans-n in #793
• Make sure to evaluate arrays before moving across threads by @lucasnewman in #798
• Add support for MOSS-TTS-Local-Transformer-v1.5 by @lucasnewman in #797
• Add Higgs v3 batch generation by @sxy-trans-n in #802
• Add Higgs v3 continuous batching by @sxy-trans-n in #804
• perf: cache ECAPA mel filters + top-k slice, remove MLX floor in sortformer by @hluaguo in #801
• perf(stt): hot-path profiling optimizations and memory/sync improvements by @hluaguo in #806
• Reduce Higgs v3 AR decode syncs by @sxy-trans-n in #808
• Point OmniVoice docs to the MLX-ready mlx-community/OmniVoice-bfloat16 by @Lazarus-931 in #807
• Higgs audio 3 stt support by @Lazarus-931 in #811
• Fix SineGen length alignment by @kanihal in #785
• Add Confucius4-TTS: multilingual zero-shot voice-cloning TTS (MLX) by @Hert4 in #799
• Add streaming and batching support for ZONOS2 by @lucasnewman in #792
• [Model] Support MOSS-Transcribe-Diarize by @gcanlin in #812
• Prevent OOM for Nemotron ASR by @lucasnewman in #817
• Add MOSS-Music by @lucasnewman in #819
• feat(stt): implement segment batching for Qwen3-ASR (--max-parallel-segments) by @hhh2210 in #783
• Bump version to v0.4.5 by @Blaizzy in #824
• Fix slow STT tests by @lucasnewman in #810
• Optimize MOSS-TD audio feature injection by @sxy-trans-n in #823

New Contributors

• @sxy-trans-n made their first contribution in #793
• @hluaguo made their first contribution in #801
• @Lazarus-931 made their first contribution in #807
• @kanihal made their first contribution in #785
• @Hert4 made their first contribution in #799
• @gcanlin made their first contribution in #812
• @hhh2210 made their first contribution in #783

Full Changelog: v0.4.4...v0.4.5