v0.6.5

What's Changed

• Consume full currency magnitude word in Zonos2 money normalizer by @selimfedakar in #1113
• Map curly double quotes in IndexTTS normalizer CHAR_MAP by @selimfedakar in #1115
• fix(qwen3_tts): apply the speech-tokenizer decoder's sliding window by @ahuzmeza in #1117
• fix(voxtral_tts): keep 25 codec context frames when streaming by @ahuzmeza in #1119
• fix(soprano): speak decimal numbers instead of a literal word.word token by @selimfedakar in #1121
• fix(longcat): stop streaming after the last chunk by @ahuzmeza in #1120
• Update MLX-Audio branding for Neywa Labs by @Blaizzy in #1123
• Allow later versions of mlx by @lucasnewman in #1128
• fix(stt): carry rounding in SRT/VTT timestamps instead of emitting SS=60 by @selimfedakar in #1124
• Nemotron ASR: add Nemotron Unified by @T0mSIlver in #1127
• Chatterbox Turbo, Higgs Audio: fix top_k=1 by @ahuzmeza in #1134
• Keep the space after a bare time in the Zonos2 normalizer by @unaltolga in #1131
• Read the cents of a dollar amount in the IndexTTS normalizer by @unaltolga in #1133
• Expand times in the Soprano text normalizer by @unaltolga in #1132
• fix(kugelaudio): pass float timesteps to the diffusion head by @ehb54 in #1130
• fix(qwen3_tts): keep frames whose first codebook id is 0 when trimming decoded audio by @HengYangDS in #1116
• Switch the Nemotron Transcribe max tokens default to the recommended value by @lucasnewman in #1135
• fix(voxcpm2): run bfloat16 checkpoints in bfloat16 by @ahuzmeza in #1141
• fix(spark): stop voice cloning from crashing by @ahuzmeza in #1140
• Split numbers from adjacent letters in the Soprano text normalizer by @unaltolga in #1139
• test(fish-drama3): Steady the decode memory tests by @stickerdaniel in #1147
• Don't end a PocketTTS sentence at a period inside a number or word by @unaltolga in #1146
• test(qwen3_tts): allow TF32 rounding in the streaming window test by @ahuzmeza in #1144
• fix: feed log-probs to make_sampler so top_p actually applies by @DimitriChrysafis in #1149
• fix(confucius4): end the text prompt with by @ahuzmeza in #1145
• fix(mimo_v2_asr): build the input embeddings in the language model's dtype by @ahuzmeza in #1143
• fix(voxtral_tts): run the flow head in the model's dtype by @ahuzmeza in #1142
• Strip curly double quotes in the LongCat text normalizer by @unaltolga in #1136
• Read zero digits after the decimal point in the MeloTTS normalizer by @unaltolga in #1137
• Bump version to 0.6.5 by @lucasnewman in #1150

New Contributors

• @T0mSIlver made their first contribution in #1127
• @unaltolga made their first contribution in #1131
• @ehb54 made their first contribution in #1130
• @HengYangDS made their first contribution in #1116
• @DimitriChrysafis made their first contribution in #1149

Full Changelog: v0.6.4...v0.6.5

v0.6.4

What's Changed

• Fix/dramabox fp32 positions by @jaddai0 in #1066
• fix(indextts): default the tokenizer to the loaded model directory by @ahuzmeza in #1078
• fix(indextts): give the mel sequence the reference's positions by @ahuzmeza in #1079
• perf(higgs_audio_v3): prefill a shared reference once per batch by @ahuzmeza in #1091
• Nemotron 3.5 Transcribe improvements by @lucasnewman in #1092
• Add Audio8-TTS-Realistic 0.6B model by @lucasnewman in #1093
• Update macOS runner to use macos-latest by @lucasnewman in #1097
• fix(whisper): hold a character split across tokens in streaming by @ahuzmeza in #1094
• fix(soprano): read ordinals as ordinals by @ahuzmeza in #1095
• Use singular currency unit for one-unit decimal amounts in expand_currency by @selimfedakar in #1103
• fix(whisper): continue streaming after the text already emitted by @ahuzmeza in #1104
• Add LFM3-Audio 2B STS model by @lucasnewman in #1107
• fix(canary): decode long audio in chunks by @ahuzmeza in #1105
• fix(fireredasr2): decode long audio in chunks by @ahuzmeza in #1106
• Keep the space after a bare time expression in expand_time by @selimfedakar in #1099
• Prime ICL streaming decoder with reference codes (fixes first-chunk timbre mismatch) by @selimfedakar in #1096
• Add NVIDIA NemotronLabs RE-USE speech enhancement by @Blaizzy in #1101
• Bump version to 0.6.4 by @lucasnewman in #1111
• fish_qwen3_omni: load quantized checkpoints without changing conversion policy by @Longfortv5 in #1108

New Contributors

• @jaddai0 made their first contribution in #1066
• @ahuzmeza made their first contribution in #1078
• @selimfedakar made their first contribution in #1103
• @Longfortv5 made their first contribution in #1108

Full Changelog: v0.6.3...v0.6.4

v0.6.3

What's Changed

• perf(audio_io): bulk PCM conversion -- 28% less round-trip time by @dexhunter in #1083
• Add Maya-2 Native TTS model by @lucasnewman in #1086
• Add Nemotron 3.5 Transcribe support and Phonon examples by @Blaizzy in #1087
• Bump version to 0.6.3 by @lucasnewman in #1088

New Contributors

• @dexhunter made their first contribution in #1083

Full Changelog: v0.6.2...v0.6.3

v0.6.2

What's Changed

• Add Ming 3.0 TTS/ASR/Audio Tiny models by @lucasnewman in #1048
• fix(fish-drama3): drop the <|semantic:0|> frame right before EOS by @ItsOdeLeo in #1052
• Flatten line breaks for Fish Audio Drama 3 to avoid early turn ends. by @lucasnewman in #1055
• fix(tts/ming): honour stream=True in Ming-3.0 generate by yielding one result per text fragment by @tlq5l in #1054
• Add Granite-Speech-5.0-2B ASR/AST model by @lucasnewman in #1056
• fix(fish-drama3): bound memory across calls and for long chunks by @stickerdaniel in #1057
• fix(tts): set the streaming chunk flags in five models by @stickerdaniel in #1058
• fix(granite_speech): load the October 2026 granite-speech-4.1-2b-plus checkpoint by @pszemraj in #1061
• Fix Gradium Phonon TTS loading and threaded generation by @Blaizzy in #1063
• Generation sampling fixes for Ming 3.0 by @lucasnewman in #1062
• fix(granite_speech): apply repetition_penalty to log-probabilities by @pszemraj in #1064
• Add Phonon 2.5 STT model support by @lucasnewman in #1065
• fix(whisper): resolve no_speech token for pre-large-v3 vocabularies by @OceanChang0406 in #1069
• Add SpeechBrain ECAPA speaker embedding extractor by @Blaizzy in #1070
• test(fish-drama3): measure decode peaks without gc noise by @stickerdaniel in #1072
• fix(vibevoice_asr): fall back to Qwen2.5 tokenizer when the local one is empty by @OceanChang0406 in #1073
• fix(fish-drama3): sample the main token without a top-k cap by @stickerdaniel in #1074
• Load Kokoro voices from the selected model snapshot by @abaichuan1017 in #1077
• fix(ming): preserve query dtype in sliding attention masks by @tokg-venv in #1075
• fix(deps): require transformers[chat-template] so chat templates can render by @tokg-venv in #1076
• Bump version to 0.6.2 by @lucasnewman in #1081

New Contributors

• @ItsOdeLeo made their first contribution in #1052
• @tlq5l made their first contribution in #1054
• @OceanChang0406 made their first contribution in #1069
• @abaichuan1017 made their first contribution in #1077
• @tokg-venv made their first contribution in #1075

Full Changelog: v0.6.1...v0.6.2

v0.6.1

What's Changed

• Ming dense: fix the prompt contract and support Ming-3.0-tts-0.5B by @cguldogan in #1047
• fix(server): use a writable temp file for STT uploads by @regcs in #1042
• fix(irodori-tts): speaker_kv_scale rollback crashes with independent CFG and dual conditioning by @matzkoh in #1044
• Add Gradium Phonon TTS model by @lucasnewman in #1049
• Bump version to 0.6.1 by @lucasnewman in #1050

New Contributors

• @matzkoh made their first contribution in #1044

Full Changelog: v0.6.0...v0.6.1

v0.6.0

What's Changed

• Fix upsampling to use the existing FIR resampler by @widechaos in #1021
• Fold Fish S1 DAC weight norm at load by @stickerdaniel in #1017
• Add Fish Audio Drama 3 by @stickerdaniel in #1018
• Optimize Breeze-TTS generation by @lucasnewman in #1010
• Fix the docs build by @lucasnewman in #1022
• fix(stt): tokenizer-safe streaming detokenization by @pszemraj in #983
• Fix Nemotron ASR CI after mlx core changes by @lucasnewman in #1032
• Fix CI for docs pages by @lucasnewman in #1033
• fish_drama3 convert: --head-bits, --expert-group-size, --expert-bits-map by @fcmeyer in #1028
• fix(server): use Qwen3-TTS generation config for omitted sampling params by @chnln in #995
• encodec, phonon: fall back to plain MLX when Metal is unavailable by @Cassian433 in #1025
• docs(fish-drama3): link an 8-bit checkpoint with bf16 dense and tower by @cguldogan in #1024
• docs(omnivoice): fix CLI flags and tokenizer path in README by @Eloitor in #1027
• fix(tts): AudioPlayer.stop() drops audio that never reached the auto-start threshold by @bennyliaw in #1001
• docs(tts): Link Drama 3 3-bit checkpoint on mlx-community by @NavSha in #1023
• fix(audio_io): decode bytes and mislabeled files, add a duration limit by @ykhrustalev in #1013
• Fix model category probing with base dependencies by @lucasnewman in #1034
• fix(stt): keep Qwen3-ASR language auto-detection per chunk by @etzelm in #1014
• Remove empty VoxCPM test by @atesahmet0 in #1035
• Add Phonon-2 speech recognition support by @atesahmet0 in #1030
• docs(qwen3_tts): Base checkpoints have no preset speakers by @hsliuustc0106 in #1000
• fix(qwen3_tts): make the 0.6B instruct guard reachable by @hsliuustc0106 in #999
• docs(qwen3_tts): clarify voice cloning vs speaker synthesis usage rules by @mm65x in #906
• Use a growing KV cache for Dia by @lucasnewman in #1037
• fix(fish_speech): prevent fast_ path and embeddings from quantizing to fix convert script by @mm65x in #905
• fix(whisper): ignore empty non-speech token encodings by @saichowdary007 in #934
• fix(stt): validate empty Qwen3-ASR audio by @ainergiz in #931
• fix(qwen3-tts): prefill full text to avoid slow speech by @atesahmet0 in #1036
• fix(stt): stop fabricating Qwen3-ASR stream timestamps by @ainergiz in #930
• fix(server): forward prompt to STT models on /v1/audio/transcriptions by @Bannng in #997
• Add MiMo-V2.5-ASR STT support by @ailuntx in #719
• feat(stt): add Granite Speech 4.1 Plus support by @pszemraj in #978
• Pin latest MLX version, fix tests by @lucasnewman in #1039
• VibeVoice performance improvements by @lucasnewman in #1040
• Use model sampling parameters by default by @lucasnewman in #1043
• Bump version to 0.6.0 by @lucasnewman in #1045

New Contributors

• @widechaos made their first contribution in #1021
• @stickerdaniel made their first contribution in #1017
• @pszemraj made their first contribution in #983
• @fcmeyer made their first contribution in #1028
• @chnln made their first contribution in #995
• @Cassian433 made their first contribution in #1025
• @cguldogan made their first contribution in #1024
• @Eloitor made their first contribution in #1027
• @bennyliaw made their first contribution in #1001
• @NavSha made their first contribution in #1023
• @atesahmet0 made their first contribution in #1035
• @hsliuustc0106 made their first contribution in #1000
• @saichowdary007 made their first contribution in #934
• @Bannng made their first contribution in #997
• @ailuntx made their first contribution in #719

Full Changelog: v0.5.9...v0.6.0

v0.5.9

What's Changed

• fix(qwen3_tts): don't add the transposed-conv bias twice in streaming decode by @Shoofio in #1003
• Add Irodori-TTS v4-Large support (T5Gemma2 text encoder) by @yoshphys in #993
• Add streaming support for Fish S2 Pro by @x-n2o in #977
• mossformer2_se: use the custom Metal kernels only when Metal is available by @joshuaswarren in #1004
• Add Nemotron speech streaming STT by @Lazarus-931 in #992
• Reuse Nemotron predictor results across blank frames by @lyonsno in #1006
• Migrate to httpx2, fix CI by @lucasnewman in #1012
• fix(server): stop capping /v1/audio/transcriptions at 1024 tokens when max_tokens is omitted by @etzelm in #1009
• Bump version to 0.5.9 by @lucasnewman in #1015

New Contributors

• @Shoofio made their first contribution in #1003
• @x-n2o made their first contribution in #977
• @joshuaswarren made their first contribution in #1004
• @etzelm made their first contribution in #1009

Full Changelog: v0.5.8...v0.5.9

v0.5.8

What's Changed

• Fix Voxtral streaming packet boundary drift by @Lazarus-931 in #990
• perf(stt): speed up Qwen3 timestamp repair by @Hosi121 in #982
• Fix Voxtral TTS openings: decode the first frame from the prompt state by @whallysson in #989
• fix(stt): preserve MOSS convolution layout when loading MLX weights by @z11i in #976
• fix(breeze): keep bf16 prompt embeddings in bf16 by @EdwardGong in #987
• Expose realtime transcription capability by @Lazarus-931 in #991
• Revert realtime transcription capability endpoint by @Lazarus-931 in #996
• Bump version to 0.5.8 by @lucasnewman in #998

New Contributors

• @Hosi121 made their first contribution in #982
• @whallysson made their first contribution in #989
• @z11i made their first contribution in #976
• @EdwardGong made their first contribution in #987

Full Changelog: v0.5.7...v0.5.8

v0.5.7

What's Changed

• Bump version to 0.5.7 by @lucasnewman in #981
• Remove errant soundfile references by @lucasnewman in #980

Full Changelog: v0.5.6...v0.5.7

v0.5.6

What's Changed

• Add Parakeet Redux by @lucasnewman in #969
• Add Nemotron 3 Diarization model by @lucasnewman in #970
• Add Nemotron ASR + diarization example by @lucasnewman in #971
• Bump version to 0.5.6 by @lucasnewman in #973

Full Changelog: v0.5.5...v0.5.6