v0.4.0

What's Changed

• add example of qwen3-asr with forced alignment by @eschmidbauer in #463
• Restore Qwen3-TTS encoder_config to preserve accents in voice clones. by @orbitalquark in #461
• Ensure TTS audio player plays a trailing, partially-filled audio frame. by @orbitalquark in #465
• Fix source separation issue with shape mismatch: noise shape for separate_long by @mnoukhov in #467
• Enable streaming for Qwen3-TTS when ICL mode is enabled. by @orbitalquark in #466
• feat(stt): add support for MedASR (Lasr architecture) by @sigjhl in #376
• Formatting fix and add to pyproject.toml by @mnoukhov in #475
• Use shared model cache resolution for SAM-Audio by @mnoukhov in #468
• Fix voice matching for Pocket TTS by @lucasnewman in #477
• Fix ALMs max tokens and chunking by @Blaizzy in #474
• [Soprano] Fix decoder and config loading (v1 and v1.1) by @Blaizzy in #480
• Add audio separation UI & Server by @Blaizzy in #347
• Add Parakeet v3 multilingual support with language detection by @andimarafioti in #481
• Revert "Do not discard the last unfilled audio frame. (#465)" by @orbitalquark in #473
• Fix longform generation for Pocket TTS by @lucasnewman in #486
• Enable streaming /v1/audio/speech server endpoint with raw/pcm data. by @orbitalquark in #484
• feat: Add support for Voxtral Mini 4B Realtime by @shreyaskarnik in #487
• fix(kokoro): Chinese TTS crashes with ValueError in g2p pipeline by @smartchainark in #489
• fix: update STT transcription parameters and preserve original audio format by @shreyaskarnik in #488
• fix(docs): update outdated model links by @joaopalmeiro in #495
• [VAD / Diarization] Add sortformer by @Blaizzy in #493
• feat: add streaming support and toggle for realtime STT by @Cold-A-Muse in #494
• chore: update protobuf dependency to version 6.33.5 by @Blaizzy in #497
• fix(ui): update footer to display the current year dynamically and link to the repo by @shreyaskarnik in #509
• Add Smart Turn v3 semantic VAD by @lucasnewman in #511
• fix(vibevoice-asr): add audio resampling and normalization to preprocessing by @bellkjtt in #510
• Refactor(whisper): update model instantiation and loading process by @Blaizzy in #514
• fix: replace 4 bare excepts with except Exception by @haosenwang1018 in #521
• feat(stt): add system_prompt parameter to Qwen3ASR generation methods by @chris-schra in #522
• fix(medasr): collapse CTC tokens manually to prevent raw output by @sigjhl in #519
• Add Echo TTS by @lucasnewman in #525
• Allow printing transcriptions to stdout when output path is "-". by @orbitalquark in #527
• stt: Keep uploaded file extension to avoid unnecessary conversions. by @orbitalquark in #528
• feat(lid): Add spoken language identification (MMS-LID) by @beshkenadze in #529
• Set max tokens to a more reasonable value by default for STT by @lucasnewman in #533
• [Qwen3-TTS] Improve inference, TTFB and add batch support by @Blaizzy in #534
• refactor(codec): extract shared ECAPA-TDNN backbone by @beshkenadze in #532
• Add KittenTTS support and ONNX parity fixes by @Reza2kn in #517
• Fix Qwen3-TTS streaming decoder throttling with incremental decoding by @Blaizzy in #537
• Add ming omni tts (MoE and Dense) by @Blaizzy in #515
• [Qwen3-ASR] Fix auto lang detection by @Blaizzy in #547
• Add nvfp4, mxfp4 and mxfp8 quants by @Blaizzy in #543
• Fix duplicate audio_samples field in GenerationResult dataclass by @Blaizzy in #548
• [Whisper] Fix lang code assignment by @Blaizzy in #549

New Contributors

• @eschmidbauer made their first contribution in #463
• @orbitalquark made their first contribution in #461
• @mnoukhov made their first contribution in #467
• @sigjhl made their first contribution in #376
• @andimarafioti made their first contribution in #481
• @shreyaskarnik made their first contribution in #487
• @smartchainark made their first contribution in #489
• @joaopalmeiro made their first contribution in #495
• @Cold-A-Muse made their first contribution in #494
• @bellkjtt made their first contribution in #510
• @haosenwang1018 made their first contribution in #521
• @chris-schra made their first contribution in #522
• @Reza2kn made their first contribution in #517

Full Changelog: v0.3.1...v0.4.0

v0.3.1

What's Changed

• Update uv.lock to reflect dependency version changes by @Blaizzy in #432
• v0.3.1: Update STT API docs and fix default output path by @Blaizzy in #433
• Qwen3-TTS: Add streaming and optimise peak usage by @Blaizzy in #435
• Fix: Use single quotes in README examples to avoid Bash history expansion. by @reinexworldc in #440
• Fix: improve import error hadling by @reinexworldc in #443
• [Qwen3-TTS] Fix some Custom Voices producing silence with 0.6B by @Blaizzy in #444
• Refactor audio load by @Blaizzy in #445
• Update pyproject.toml for poetry support by @lucasnewman in #446
• Add Qwen3-ASR by @Blaizzy in #454
• Fix chatterbox load by @Blaizzy in #455
• Update README to remove basic usage section by @Blaizzy in #456
• Update README with output path for ASR commands by @rahimnathwani in #458
• Update package dependencies in uv.lock to include new extras by @Blaizzy in #457
• Fix server (STT, TTS) by @Blaizzy in #460

New Contributors

• @reinexworldc made their first contribution in #440

Full Changelog: v0.3.0...v0.3.1

v0.3.0

What's Changed

• Fix speaker embedding extraction in Qwen3-TTS model by @Blaizzy in #390
• Fix Qwen3-TTS tail artifacts by @Blaizzy in #391
• Fix Qwen3-TTS Base Voice Cloning by @Blaizzy in #394
• Add Vibevoice ASR by @Blaizzy in #389
• Qwen3 speaker embedding tests by @Blaizzy in #396
• Update TTS commands in README to include language code option by @rudolfolah in #401
• Unify Mimi implementation for Pocket TTS by @lucasnewman in #403
• Fix issue of ref_audio not loading prior to inference with server. by @BuffMcBigHuge in #406
• Enhance README with installation and usage examples by @rahimnathwani in #404
• Upgrade GitHub Actions for Node 24 compatibility by @salmanmkc in #418
• Upgrade GitHub Actions to latest versions by @salmanmkc in #419
• [VibeVoice-ASR] Fix Metal kernel crash and optimize memory for long audio by @Blaizzy in #417
• fix: Allowing quantization of Qwen3-TTS! Adding model_quant_predicate to Qwen3-TTS to exclude embedding layers by @kyr0 in #398
• Fix qwen3 tts quants (silence in VC and word precision) by @Blaizzy in #407
• Fix stt array io by @Blaizzy in #426
• Update MANIFEST.in to remove leading dot from requirements.txt path by @Blaizzy in #428
• Move audio path/format prints under verbose flag by @wladpaiva in #429
• Update pyproject.toml and GitHub Actions workflow for package publishing by @Blaizzy in #431

New Contributors

• @rudolfolah made their first contribution in #401
• @BuffMcBigHuge made their first contribution in #406
• @rahimnathwani made their first contribution in #404
• @salmanmkc made their first contribution in #418
• @kyr0 made their first contribution in #398
• @wladpaiva made their first contribution in #429

Full Changelog: v0.2.10...v0.3.0

v0.3.0rc1

What's Changed

• Remove extra deps by @Blaizzy in #373
• Refactor load by @Blaizzy in #374
• Add lfm2 audio by @Blaizzy in #370
• update lfm readme by @Blaizzy in #377
• Fix lang codes kokoro by @Blaizzy in #380
• Replace soundfile with miniaudio + ffmpeg by @Blaizzy in #379
• Add Pocket TTS model by @lucasnewman in #381
• Fix STT stream by @Blaizzy in #382
• Migrate swift to https://github.com/Blaizzy/mlx-audio-swift by @Blaizzy in #363
• Refactor model path retrieval in get_model_path function by @Blaizzy in #383
• Add streaming decoding to snac and orpheus by @Blaizzy in #384
• Update generate output path by @Blaizzy in #385
• Add Qwen3-TTS by @Blaizzy in #388

Full Changelog: v0.2.10...v0.3.0rc1

v0.2.10

What's Changed

• Refactor GLMASR and improve LM style ASR logging by @Blaizzy in #332
• Remove actual issue ID reference from PR template by @mootari in #334
• Add maya1 fixes to Llama by @Blaizzy in #340
• Fix marvis, chatterbox and args by @Blaizzy in #342
• Add Sam Audio by @Blaizzy in #338
• fix: Add missing mlx-lm dependency by @joshwhiton in #344
• feat(swift): add Kokoro-82M-v1.1-zh MLX Support by @Alex-Wengg in #341
• Remove loguru reconfiguration on Kokoro import by @joshwhiton in #348
• feat(stt): add AlignAtt streaming transcription for Whisper by @beshkenadze in #321
• Fix stft args by @Blaizzy in #354
• Allow using DACVAE as a codec independent of SAM Audio model by @lucasnewman in #357
• chore: update Python version requirement and dependencies by @Blaizzy in #355
• Add MossFormer2 SE (Speech Enhancement) by @starkdmi in #351
• use chatterbox MTLTokenizer for multilingual. by @litmudoc in #362
• Add streaming and refactor Sam Audio API by @Blaizzy in #360
• Add Soprano by @Blaizzy in #359
• Fix model type, refactor orpheus style models by @Blaizzy in #358
• revert default response format to mp3 by @Blaizzy in #356
• Refactor voice loading in KokoroPipeline to support .safetensors files by @Blaizzy in #364
• Add uv.lock and pin all deps as core by @Blaizzy in #366

New Contributors

• @mootari made their first contribution in #334
• @Alex-Wengg made their first contribution in #341
• @starkdmi made their first contribution in #351
• @litmudoc made their first contribution in #362

Full Changelog: v0.2.9...v0.2.10

v0.2.9

What's Changed

• Add GLM ASR by @Blaizzy in #320
• Simplify convert API for TTS and STT by @Blaizzy in #324
• [Chatterbox-turbo] Add speaker embedding by @Blaizzy in #322
• [Chatterbox-turbo] Add in-place cache by @Blaizzy in #322
• [Chatterbox-turbo] Add audio streaming by @Blaizzy in #322
• [Chatterbox-turbo] Add audio chunking by @Blaizzy in #322

Full Changelog: v0.2.6...v0.2.9

v0.2.8

What's Changed

• fix(server): use lowercase for default response_format by @beshkenadze in #301
• Add Chatterbox and Chatterbox Turbo by @Blaizzy in #302
• Add Chatterbox [VC only] by @DePasqualeOrg in #282
• feat: add lazy imports for TTS/STT modules by @beshkenadze in #290
• Pin tfms dep <5.0.0 by @Blaizzy in #303
• feat: migrate from setup.py to pyproject.toml with optional deps by @beshkenadze in #291
• fix(test): use case-insensitive content-type comparison by @beshkenadze in #300
• ci: add modular installation tests for pyproject.toml extras by @beshkenadze in #298
• Fix build by @Blaizzy in #304

Full Changelog: v0.2.7...v0.2.8

v0.2.7

What's Changed

• Refactor Marvis TTS API: Make public methods accessible by @rudrankriyam in #259
• Add Marvis model selection to TTS UI by @rudrankriyam in #261
• Add Marvis quant selection to TTS Web UI by @adrgrondin in #264
• Fix security vulnerabilities in Next.js and brace-expansion dependencies by @Copilot in #265
• make the methods more useful by @pritamsoni-hsr in #246
• Update to mlx-swift-lm and remove redundant mlx-swift dependency by @rudrankriyam in #269
• Fix voxtral segments by @Blaizzy in #273
• Add UI startup option to Server by @Blaizzy in #274
• Add preemphasis preprocessing support for Parakeet models to match NeMo training config by @joshwhiton in #286
• feat: Add support for VoxCPM (w/ voice cloning) by @voxmenthe in #293
• Fix spark decoding by @Blaizzy in #296
• feat: extract DSP utilities to dedicated module by @beshkenadze in #289
• Feat: add response format option to SpeechRequest by @Blaizzy in #297
• Add VibeVoice by @Blaizzy in #295

New Contributors

• @pritamsoni-hsr made their first contribution in #246
• @joshwhiton made their first contribution in #286
• @voxmenthe made their first contribution in #293
• @beshkenadze made their first contribution in #289

Full Changelog: v0.2.6...v0.2.7

v0.2.6

What's Changed

• fix wav2vec by @josharian in #222
• Fix RTF calculation in kokoro model by @davidxifeng in #227
• Fix Unnecessary Audio Transcription for the IndexTTS Model by @bytefer in #231
• Add Sesame TTS Integration for Swift Audio Package by @rudrankriyam in #223
• Use batched vocoding to reduce peak memory usage with Sesame arch models by @lucasnewman in #236
• Cache RoPE by dtype for Sesame arch models for improved generation performance by @lucasnewman in #232
• Install Metal toolchain for Swift tests by @lucasnewman in #233
• Adopt changes interface changes from mlx-lm to fix Sesame-arch models by @lucasnewman in #242
• Update swift-transformers dependency to 1.1.0 by @Liam1506 in #247
• Improve Swift TTS app UX by @rudrankriyam in #248
• Add quality selection and streaming controls to Marvis with UI support for macOS & iOS by @rudrankriyam in #249
• Fix Swift compiler warnings by @rudrankriyam in #250
• Refactor MarvisModel to handle optional backbone and decoder flavors by @rudrankriyam in #251
• Fix iOS 16 compatibility and ESpeakNG framework linking for iOS app by @rudrankriyam in #252
• Add memory increase limit for iOS by @rudrankriyam in #253
• Update audio playback management in Marvis TTS by @rudrankriyam in #254
• Bump version and add new copy files by @Blaizzy in #255
• Add UI v2 by @Blaizzy in #154

New Contributors

• @josharian made their first contribution in #222
• @davidxifeng made their first contribution in #227
• @bytefer made their first contribution in #231
• @Liam1506 made their first contribution in #247

Full Changelog: v0.2.5...v0.2.6

v0.2.5

What's Changed

• Use indeterminate progress for CSM models by @lucasnewman in #216
• Bump version to 0.2.5 by @Blaizzy in #219

Full Changelog: v0.2.4...v0.2.5