v0.1.8

What's Changed

• Add VoxCPM TTS model by @lucasnewman in #319
• Add Phonon-2.5 models by @lucasnewman in #314
• Add OuteTTS model by @lucasnewman in #320
• Qwen3-TTS: keep the ICL prompt in the talker's dtype by @ahuzmeza in #315
• Cache packages to speed up CI by @lucasnewman in #323
• Add Higgs Audio v2 TTS model by @lucasnewman in #322
• Add rumik-oss TTS model by @lucasnewman in #321
• Add Higgs Audio v3 TTS model by @lucasnewman in #324
• Summarize models in README by @lucasnewman in #325
• Add Bark TTS model by @lucasnewman in #326
• Add Audio8 TTS model by @lucasnewman in #327
• Add VibeVoice Realtime model by @lucasnewman in #328
• Add KugelAudio model by @lucasnewman in #329
• Add Ming TTS model by @lucasnewman in #330
• Add Ming-3.0 TTS Tiny model by @lucasnewman in #331
• Add VibeVoice TTS model by @lucasnewman in #332
• IndexTTS: give the mel sequence the reference's positions by @ahuzmeza in #316
• Add LongCat TTS model by @lucasnewman in #337
• Add ZONOS2 TTS model by @lucasnewman in #338
• Add Audio8 hybrid TTS model by @lucasnewman in #339
• Add MeloTTS model by @lucasnewman in #340
• Add Voxtral TTS by @lucasnewman in #342
• Add Confucius4-TTS model by @lucasnewman in #341
• Add TADA TTS model by @lucasnewman in #344
• Add Ming Omni TTS model by @lucasnewman in #343
• Whisper: fix words lost at the 30 s cuts in long audio by @ahuzmeza in #345
• Parakeet: fix streaming repeating the whole transcript after a chunk merge by @ahuzmeza in #346
• Add OmniVoice 8-bit and voice-cloning helpers by @petik3 in #335
• Add Nemotron 3.5 Transcribe STT model by @lucasnewman in #348
• Fix blocking behavior when streaming with STT models by @lucasnewman in #349
• Qwen3-TTS: fix the upsample bias counted twice at chunk joins by @ahuzmeza in #350
• Updates for mlx-swift 0.33.0 by @lucasnewman in #359
• Soprano: fix ordinals read as "trillion" by @ahuzmeza in #358
• Mimi: fix attention seeing past its context window (PocketTTS cloning drops the first word) by @ahuzmeza in #357
• loadAudioArray: fix stereo files keeping only the left channel by @ahuzmeza in #355
• VoxCPM2: fix Chinese text coming out garbled by @ahuzmeza in #356
• Adopt MaterializedArray by @lucasnewman in #360
• Whisper, Cohere Transcribe: fix streaming resending the whole transcript on non-Latin text by @ahuzmeza in #351
• Qwen3-TTS: fix reference audio ignored when no transcript is given by @ahuzmeza in #352
• Kokoro, KittenTTS: fix English numbers read wrong or skipped by @ahuzmeza in #353
• Kokoro: fix accented letters in English text being dropped by @ahuzmeza in #354
• Canary: fix language names dropping the language tokens by @ahuzmeza in #361
• Mel spectrogram: drop the extra last frame, as Python does by @T0mSIlver in #363
• FireRedASR2: fix long audio dropping sentences and looping by @ahuzmeza in #362
• GLM-ASR: fix audio after 30 s being dropped by @ahuzmeza in #365
• Canary: fix long audio drifting and stopping early by @ahuzmeza in #364
• Add missing MLXFFT dependency to MLXAudioSTT by @ahuzmeza in #366
• Marvis, Mimi: fix the crash on reference audio of some lengths by @ahuzmeza in #367
• Mimi: fix resetState keeping the end of the previous clip by @ahuzmeza in #368

New Contributors

• @ahuzmeza made their first contribution in #315
• @petik3 made their first contribution in #335

Full Changelog: v0.1.7...v0.1.8

v0.1.7

What's Changed

• VoxtralRealtime streaming encoder: write KV rows into preallocated storage by @T0mSIlver in #272
• Nemotron streaming: compute only new mel frames; keep MLX's buffer cache between steps by @T0mSIlver in #274
• VoxtralRealtime streaming: reuse the mel Hann window instead of rebuilding it per append by @T0mSIlver in #271
• VoxtralRealtime generateStream: build deltas from token bytes by @T0mSIlver in #273
• Fix speaking rate in non-streaming Qwen3-TTS generation by @lucasnewman in #290
• Irodori-TTS: add v4.1 and v4.1-MF (MeanFlow) support by @jonpol01 in #284
• Fix partial model downloads returning as valid from the HF cache by @lucasnewman in #291
• Nemotron streaming: reuse decoder and encoder work that has not changed by @T0mSIlver in #275
• Build against mlx-swift main (#469): Parakeet and Qwen3-TTS compile calls by @T0mSIlver in #292
• STT: forward the loadModel cache to the six loaders that dropped it by @T0mSIlver in #293
• Qwen3-ASR: cap maxTokens per chunk, not per file by @T0mSIlver in #294
• Add Mega-ASR model by @lucasnewman in #299
• Add Phonon 2 STT model by @lucasnewman in #297
• Whisper: add a translate task to STTGenerateParameters by @T0mSIlver in #295
• Add Granite Speech TurboCTC STT model by @lucasnewman in #298
• Add Fun-ASR-Nano STT model by @lucasnewman in #300
• Add Higgs Audio v3 STT model by @lucasnewman in #301
• Add Parakeet Redux STT model by @lucasnewman in #302
• Core: reuse a cached snapshot that has no config.json by @T0mSIlver in #296
• Add Granite Speech 4.1 and 5.0 STT models by @lucasnewman in #303
• Add Nemotron-VoiceChat codec by @lucasnewman in #304
• Add MiMo codec by @lucasnewman in #305
• TTS: Kokoro and Kitten text processors use the loadModel cache by @T0mSIlver in #307
• Add Qwen2-Audio model by @lucasnewman in #306
• Add offline Voxtral Mini support by @lucasnewman in #308
• Add VibeVoice ASR support by @lucasnewman in #309
• Add MiMo ASR model by @lucasnewman in #311
• Add Ming 3.0 ASR Tiny model by @lucasnewman in #312
• Add MOSS-Music model by @lucasnewman in #313

Full Changelog: v0.1.6...v0.1.7

v0.1.6

What's Changed

• support for dia by @Lazarus-931 in #262
• Add VoxCPM2 TTS model (2B params, 48kHz, voice cloning) by @sb1992 in #175
• Fix parallel test runner by @lucasnewman in #282
• Remove dependencies on MLXLMCommon / MLXLLM by @lucasnewman in #287
• Add Gradium Phonon TTS model by @lucasnewman in #288
• Add translation example by @lucasnewman in #289

New Contributors

• @sb1992 made their first contribution in #175

Full Changelog: v0.1.5...v0.1.6

v0.1.5

What's Changed

• Update for mlx-swift / mlx-swift-lm by @lucasnewman in #281

Full Changelog: v0.1.4...v0.1.5

v0.1.4

What's Changed

• Fix float32 leak in VoxtralRealtime streaming by @T0mSIlver in #226
• VoxtralRealtime streaming: incremental mel/conv front end (O(N²) → O(N) per utterance) by @T0mSIlver in #230
• VoxtralRealtime streaming: stop clearing the Metal buffer pool on every step by @T0mSIlver in #229
• Route Qwen3-ASR text attention through attentionWithCacheUpdate by @iannellomarco in #228
• Hoist per-layer-invariant attention inputs out of the Voxtral layer loops by @T0mSIlver in #231
• Honor Chatterbox emotion override with default conditioning by @yanchenko in #234
• VoxtralRealtime: load checkpoints with a quantized tied embedding by @T0mSIlver in #232
• Quantized KV cache for MOSS-Transcribe-Diarize (opt-in, memory-bounded long-form) by @iannellomarco in #225
• Add reproducible OmniVoice generation seed by @yanchenko in #233
• OmniVoice: honor task cancellation during generation by @raghakot in #227
• Stream Fish Speech progressively by @steipete in #237
• Support English Nemotron streaming checkpoints by @obra in #236
• Fix quantized Whisper checkpoint loading by @Parassharmaa in #235
• OmniVoice: emit per-step denoise progress through generateStream by @raghakot in #219
• Kokoro: expose the per-phoneme durations generate already computes by @SAPSentinel in #239
• fix(package): exclude SileroVAD README from the MLXAudioVAD target by @Pietroski in #250
• FireRedASR2: fix beam search topK gathering across rows by @Kn0688 in #253
• STT: scale fbank input to int16 unconditionally in FireRedASR2/SenseVoice by @Kn0688 in #256
• feat: add Breeze TTS 2 support by @dychen1 in #255
• Fix Qwen3-ASR mel frontend: Slaney mel scale + periodic Hann window by @ZigZagYea in #247
• support for spark-tts by @Lazarus-931 in #261
• Fix Spark tests in CI by @lucasnewman in #268
• VoxtralRealtime streaming: drop conv and adapter rows once they are consumed by @T0mSIlver in #263
• VoxtralRealtime decoder: append KV rows in place instead of rebuilding the window by @T0mSIlver in #264
• VoxtralRealtime streaming: decode the transcript one token at a time by @T0mSIlver in #265
• Improve padding for Qwen3-ASR batched inference by @lucasnewman in #269
• Fix test failures and remove trivial tests by @lucasnewman in #276

New Contributors

• @T0mSIlver made their first contribution in #226
• @iannellomarco made their first contribution in #228
• @yanchenko made their first contribution in #234
• @raghakot made their first contribution in #227
• @steipete made their first contribution in #237
• @obra made their first contribution in #236
• @Parassharmaa made their first contribution in #235
• @SAPSentinel made their first contribution in #239
• @Pietroski made their first contribution in #250
• @Kn0688 made their first contribution in #253
• @dychen1 made their first contribution in #255
• @ZigZagYea made their first contribution in #247
• @Lazarus-931 made their first contribution in #261

Full Changelog: v0.1.3...v0.1.4

v0.1.3

What's Changed

• Fill in some missing codecs by @lucasnewman in #96
• Add Echo TTS by @lucasnewman in #97
• Add FireRed ASR 2 by @lucasnewman in #98
• Add SenseVoice (ASR/AER/LID) by @lucasnewman in #99
• Add Chatterbox Turbo by @lucasnewman in #102
• Update README.md by @Blaizzy in #104
• Add Fish Audio S2 Pro model by @lucasnewman in #106
• Add README for Qwen3 TTS by @lucasnewman in #107
• Fix Parakeet multilingual recognition (Russian/non-English) by @beshkenadze in #108
• Allow autodetecting language for Qwen3 ASR by @lucasnewman in #110
• feat: add MLXAudioG2P — neural ByT5 + dictionary lexicons by @beshkenadze in #121
• feat: add StyleTTS2 shared blocks, ALBERT encoder, and English G2P by @beshkenadze in #122
• feat: add KittenTTS text-to-speech model by @beshkenadze in #123
• feat: add Kokoro TTS with multilingual support by @beshkenadze in #124
• Add MLXFast dependency for MLXAudioCodecs by @noorbhatia in #128
• Add missing MLXFast dependency to MLXAudioSTT by @Hurryitup in #132
• Adding fromModelDirectory methods to various models by @ptsochantaris in #144
• fix: Kokoro 8-bit crash on quantized weight layout and NaN durations by @beshkenadze in #136
• Add Cohere Transcribe STT support by @beshkenadze in #129
• fix: Parakeet int8 Compatibility Fix by @RaihanulHaque in #145
• fix: Qwen3 TTS - handle mixed types in spkId and spkIsDialect for CustomVoice model by @Ichigo3766 in #146
• feat: speaker embedding support for Qwen3-TTS CustomVoice models by @Ichigo3766 in #148
• fix(cohere): load unmodified Python-produced quantized checkpoints by @beshkenadze in #153
• [Qwen3 TTS] Cache reference audio between generations by @lucasnewman in #125
• Add context support to Qwen3 ASR by @lucasnewman in #126
• fix: pass textProcessor through in KokoroModel.fromPretrained by @Oliver2213 in #151
• fix: exclude in-tree docs from SwiftPM targets by @gaelic-ghost in #152
• models: hide marvis rope caches from strict verification by @gaelic-ghost in #163
• models: escape underscore member reads in EnglishG2P by @gaelic-ghost in #161
• Upgrade package versions and adopt changes from mlx-lm-swift by @lucasnewman in #158
• Add DeepFilterNet speech enhancement model by @kylehowells in #100
• Parakeet(perf): batched decode, bf16 compute, TDT perf fixes by @beshkenadze in #167
• Add optional progressHandler to ModelUtils.resolveOrDownloadModel by @Oliver2213 in #166
• feat: add reusable qwen3 reference conditioning by @gaelic-ghost in #156
• docs: add contributing guide by @beshkenadze in #142
• Fix (Qwen3TTS): pass topK and minP from GenerateParameters by @Oliver2213 in #168
• Improved task cancellation handling for generation by @lucasnewman in #157
• Fix Qwen3ASR stream cancellation propagation by @summereasy in #172
• Add MOSS-TTS-Nano model by @lucasnewman in #169
• Add Silero VAD model by @beshkenadze in #176
• Qwen3-ASR: chunked prefill + asyncEval + repetition penalty + loop guard by @beshkenadze in #174
• Add MOSS-TTS model family by @lucasnewman in #179
• fix(MossTTS): gate homeDirectoryForCurrentUser behind #if os(macOS) by @brunobar79 in #189
• FireRedASR2Model: hide CMVN runtime arrays from strict-verify reflection by @heyzgj in #188
• perf(sortformer): single bulk readback in predsToSegments (~1.8x streaming) by @beshkenadze in #193
• Add Nemotron 3.5 ASR streaming model support by @vanch007 in #195
• Nemotron 3.5 ASR: fix prompt_kernel load crash + add cache-aware streaming (builds on #195) by @beshkenadze in #196
• refactor(stt): extract shared NeMo-family module; decouple NemotronASR from Parakeet (#197) by @beshkenadze in #198
• Kokoro: fix French nasal vowels dropped by Character tokenization by @slikmedia in #204
• Add Irodori-TTS (Japanese flow-matching TTS) by @jonpol01 in #206
• Add Whisper model family by @JacobLinCool in #192
• feat(voxtral): true incremental streaming for Voxtral Realtime by @beshkenadze in #205
• Fix Qwen3-TTS CustomVoice voice parsing by @summereasy in #186
• feat(nemotron_asr): incremental NemotronASRStreamSession for live mic by @beshkenadze in #208
• Voxtral Realtime: add Silero VAD pre-processor by @beshkenadze in #178
• Cohere Transcribe: add Silero VAD pre-processor by @beshkenadze in #177
• Add OmniVoice TTS (NAR diffusion over Qwen3 + HiggsAudioV2 codec, with voice cloning) by @JacobLinCool in #209
• docs(omnivoice): bf16 weights now include the complete tokenizer by @beshkenadze in #213
• Fix remote tokenizer loading for the MOSS models by @lucasnewman in #207
• Add Xcode project gen for SimpleChat example by @lucasnewman in #211
• Fill in remaining codecs for parity with mlx-audio by @lucasnewman in #214
• Add additional STT models by @lucasnewman in #215
• Add MOSS-Transcribe-Diarize by @lucasnewman in #221
• Add IndexTTS by @lucasnewman in #220
• Improve Voices STT controls and MOSS streaming by @Blaizzy in #223

New Contributors

• @noorbhatia made their first contribution in #128
• @Hurryitup made their first contribution in #132
• @ptsochantaris made their first contribution in #144
• @RaihanulHaque made their first contribution in #145
• @Ichigo3766 made their first contribution in #146
• @Oliver2213 made their first contribution in #151
• @gaelic-ghost made their first contribution in #152
• @kylehowells made their first contribution in #100
• @summereasy made their first contribution in #172
• @brunobar79 made their first contribution in #189
• @heyzgj made their first contribution in #188
• @vanch007 made their first contribution in #195
• @slikmedia made their first contribution in #204
• @jonpol01 made their first contribution in #206
• @JacobLinCool made their first contribution in #192

Full Changelog: v0.1.2...v0.1.3

v0.1.2

What's Changed

• feat: add MLXAudioLID module for spoken language identification by @beshkenadze in #80
• fix Parakeet V2 mel scale and add eval for joint output by @ekryski in #86
• Fix window chunking for Qwen3 ASR by @lucasnewman in #92
• Add clarifying comments on how to run tests, and fix test artifact remnants by @lucasnewman in #94
• Add granite speech 4 by @Blaizzy in #95

New Contributors

• @ekryski made their first contribution in #86

Full Changelog: v0.1.1...v0.1.2

v0.1.1

What's Changed

• Use an OSS model instead of the on-device model for the SimpleChat example by @lucasnewman in #71
• Update README.md to reflect new audio models and enhancements by @Blaizzy in #72
• Add a CLI option to emit timestamps for speech <-> text alignment by @lucasnewman in #74
• Enhance model download validation and cache management in ModelUtils by @Blaizzy in #76
• Add an MLXArray-based audio resampling variant by @lucasnewman in #77
• VoicesApp - STS view added by @alpaycli in #78
• Improve performance of Qwen3 TTS by @lucasnewman in #81
• More performance improvements for Qwen3 TTS by @lucasnewman in #82

New Contributors

• @alpaycli made their first contribution in #78

Full Changelog: v0.1.0...v0.1.1

v0.1.0

What's Changed

• Add DACVAE Codec by @Blaizzy in #2
• Add CI/CD tests by @Blaizzy in #3
• Add GLM ASR (STT) by @Blaizzy in #4
• Stop audio and generation when backgrounded to fix app crash by @adrgrondin in #8
• Fix the framework build so it can be referenced in Xcode by @lucasnewman in #10
• MLX-Audio Swift SDK v1 by @Blaizzy in #1
• Update README for modular SDK architecture and remove Kokoro TTS by @Copilot in #11
• Project cleanup by @lucasnewman in #12
• Add a basic CLI tool for TTS generation by @lucasnewman in #15
• Add support for Marvis TTS by @lucasnewman in #16
• Add Pocket TTS by @lucasnewman in #19
• Easier Voices app build by @lucasnewman in #20
• Fix voice matching for Pocket TTS by @lucasnewman in #21
• Use voices from the model repo instead of hardcoding by @lucasnewman in #22
• Move STT models to Models folder by @Blaizzy in #24
• Add Qwen3 ASR by @Blaizzy in #26
• Fix Soprano model by @lucasnewman in #28
• Add READMEs for each STT/TTS model by @lucasnewman in #27
• Add a simple local chat example by @lucasnewman in #25
• [Qwen3-ASR] Add live transcript support and lower peak usage by @Blaizzy in #32
• Add Qwen3-TTS conditional generation (VoiceDesign) support by @INQTR in #23
• [VAD / Diarization] Add sortformer by @Blaizzy in #33
• Refactor tests by @Blaizzy in #34
• Improved TTS sampling param handling by @lucasnewman in #30
• [STS] Add MossFormer2 SE speech enhancement module by @beshkenadze in #29
• [Qwen3ASR] Add STT + Live transcript to UI by @Blaizzy in #39
• Add remaining support for Qwen3 TTS base models by @lucasnewman in #45
• fix(sts): MossFormer2 speech enhancement quality — Kaldi fbank, RoPE, local loading by @beshkenadze in #44
• Add SAM Audio STS model by @lucasnewman in #46
• Add command line tool for STT by @lucasnewman in #47
• Fix loading of quantized Soprano models by @lucasnewman in #50
• Add Parakeet STT models by @lucasnewman in #51
• Add Voxtral Realtime STT model by @lucasnewman in #52
• Add LFM-2.5-Audio by @Blaizzy in #53
• Refactor sts app by @Blaizzy in #56
• Fix quantization for Qwen3 TTS models by @lucasnewman in #61
• Fix Voices App Example by @ikenwoo in #59
• Allow configuring the cache location for downloaded models by @lucasnewman in #60
• Remove dead code and fix all warnings by @lucasnewman in #62
• Add streaming support to the built-in audio player by @lucasnewman in #63
• Add Smart Turn v3 model by @lucasnewman in #64
• Fix Swift 6.2 compilation errors: @sendable closures and OptionSet type inference by @BenRacicot in #66
• Semantic VAD example by @lucasnewman in #67
• Add stock voices and clean up UI in the Voices app by @lucasnewman in #68
• Fix audio entitlements for MacOS and file validation in resolveDownload by @Blaizzy in #69
• Add streaming audio chunk support in Qwen3TTS model by @Blaizzy in #70

New Contributors

• @Blaizzy made their first contribution in #2
• @adrgrondin made their first contribution in #8
• @lucasnewman made their first contribution in #10
• @Copilot made their first contribution in #11
• @INQTR made their first contribution in #23
• @beshkenadze made their first contribution in #29
• @ikenwoo made their first contribution in #59
• @BenRacicot made their first contribution in #66

Full Changelog: https://github.com/Blaizzy/mlx-audio-swift/commits/v0.1.0