v0.9.4

What's Changed

• UI: add recipe-based cross-language dubbing and fix Studio/Arena state by @0xShug0 in #1106
• Echo-TTS: Sway time scheduling by @lojack5 in #1105
• Add opt-in VoxCPM2 ultimate cloning by @ersensari in #982
• Add Music Flamingo and SpotSound LoRA support by @0xShug0 in #1107
• feat(webui): allow multiple model package downloads, Fix #1075 by @ZeroDot1 in #1082
• omnivoice: size automatic long-form chunks in code points, not bytes by @linyqh in #1108
• OmniVoice: match Python duration weights for Bengali marks, danda and digits by @kawshikbuet17 in #1110
• Speed up Earshot VAD with per-ISA SIMD kernels by @Etherll in #1102
• OmniVoice: apply the tokenizer's NFC normalizer to input text by @kawshikbuet17 in #1116
• Fix Parakeet offline chunking for stereo and non-16-kHz audio by @0xShug0 in #1117
• Add --list-devices --json with PCI ids and device memory by @Etherll in #1118
• Fix Vulkan RMSNorm dispatch overflow for long voice references by @0xShug0 in #1120
• Add Magenta RealTime 2 Small and Base with streaming generation by @0xShug0 in #1121
• feat: add lfm3_audio for LFM3-Audio-2B ASR, TTS and S2S by @ykhrustalev in #1119
• fix: return 400 for malformed option values by @ykhrustalev in #1123
• build(nix): use stdenv.hostPlatform for platform checks by @luizribeiro in #1125
• Restore ordered model loader selection by @0xShug0 in #1126
• Move package loading rules into model-owned resolvers by @0xShug0 in #1127
• Fix Arena package selection with restored Studio settings by @0xShug0 in #1128
• new asr model: parakeet_tdt: add Underdog Bark 3.0 Pro English ASR by @mirek190 in #1130
• Add GGML implementation selection and gfx1151_v1 HIP kernels by @0xShug0 in #1134
• fix: bound the resample kernel cache by memory with LRU eviction by @ykhrustalev in #1124
• new asr model: parakeet_tdt: add Underdog Bark 3.0 Plus English ASR by @mirek190 in #1131
• fix: keep lfm2_audio and lfm3_audio fixed ASR chunks within the cap by @ykhrustalev in #1135
• OmniVoice: make the audio tokenizer codebook weight type configurable by @kawshikbuet17 in #1129
• perf: reuse lfm3_audio backbone head and prefill block graphs by @ykhrustalev in #1136
• echo-tts: minor docs cleanup (#1105 followup) by @lojack5 in #1140
• Expose backend implementation selection (rollout is deferred) by @0xShug0 in #1138
• new asr model: parakeet_tdt: add standard Underdog Bark 3.0 English ASR by @mirek190 in #1132
• Preserve backend implementation selection in Qwen3-TTS by @0xShug0 in #1141
• Allow gfx1151_v1 kernel selection on other RDNA GPUs by @0xShug0 in #1142
• fix: fetch BoringSSL with CMake older than 3.24 by @ykhrustalev in #1144
• Load an installed package by id through POST /v1/models/load by @HamSanders in #1098
• Fix OmniVoice instruct splitting for UTF-8 separators by @supomarlo in #1146
• cuda: gate Flash Attention zero-fill workaround to NVCC 13.4 by @askmyteapot in #1143
• Extend HIP batched F32 matmul and IM2COL kernels by @0xShug0 in #1150
• Use Flash Attention for RoFormer separation on Vulkan by @ruihe774 in #1152
• fix: scale Qwen3 forced aligner timestamps to the input sample rate (#1147) by @m-abs in #1148
• community: add Nemotron 3.5 Transcribe 30B-A4B by @JarodMica in #1137
• Add Nemotron Transcribe CPU expert offload and reduce resident memory by @0xShug0 in #1154
• Minor Parakeet graph and resource lifecycle cleanup by @0xShug0 in #1155
• Expose Qwen3-ASR performance controls and cuda_perf preset by @0xShug0 in #1156
• Optimize Parakeet subsampling on Metal by @0xShug0 in #1157
• Magenta RealTime 2: accept the legacy duration_seconds field and add the missing WebUI catalog entry by @noctrex in #1158
• Update WebGPU backend and add audio operator support by @0xShug0 in #1159
• Add native WebGPU backend integration by @0xShug0 in #1160
• Rename Whistle decoder exec_plan implementation to native by @0xShug0 in #1161
• Add browser WebGPU support by @0xShug0 in #1162
• feat(irodori_tts): support Irodori-TTS v4 Large (T5Gemma 2 text encoder) by @KKTTSPJ in #1163
• Extend video VAE runtimes and opt-in CUDA lowerings by @0xShug0 in #1165
• fix(irodori_tts): normalize input text like Python normalize_text by @KKTTSPJ in #1164
• Add opt-in artifact output to generic task streams by @0xShug0 in #1166
• Add SoulX-FlashHead Lite streaming portrait generation by @0xShug0 in #1167
• perf(irodori_tts): read the RF context K/V cache in place by @KKTTSPJ in #1168
• perf(irodori_tts): leave caption tokens out of the RF context without a caption by @KKTTSPJ in #1170
• fix(metal): correct SSM scan token-block reduction indexing by @0xShug0 in #1171
• ggml: add Metal IM2COL_3D support by @0xShug0 in #1169
• ggml: add WebGPU GROUP_NORM support by @0xShug0 in #1172
• perf(metal): optimize SSM scan and convolution by @0xShug0 in #1173
• fix(torch_bin): read checkpoints that wrap the state dict and hold lists by @KKTTSPJ in #1174
• feat(ui): add streaming voice chat with LFM3 and SoulX avatar by @0xShug0 in #1176
• Improve Metal SSM performance by @0xShug0 in #1179
• Synchronize shared storage in Mamba-1 scan by @0xShug0 in #1181
• Separate model runtime tests from framework unit tests by @0xShug0 in #1182
• Correct Vulkan pipeline selection for oversized buffer bindings by @0xShug0 in #1183
• Clean up first-party compiler warnings (WIP) by @0xShug0 in #1184
• Separate Parakeet, Phonon-2, and Bark ASR family routing by @0xShug0 in #1189
• Find hipcub and rocprim in the HIP build by @apt-install-coffee in #1186
• Add Haas Japanese ASR using the shared Parakeet runtime by @0xShug0 in #1191
• Add Real-time RE-USE with incremental audio streaming by @0xShug0 in #1192
• nemotron_transcribe: normalize regional language hints by @JarodMica in #1190
• Add MuScriptor Medium and Large F32 variants by @0xShug0 in #1193
• Add PocketTTS Small with streaming and acoustic batching by @0xShug0 in #1194
• Add WebUI voice library export, import, and sorting by @0xShug0 in #1195
• Keep HIP release pip cache on temporary scratch storage by @0xShug0 in #1196
• Add an opt-in repetition guard for Nemotron Transcribe by @0xShug0 in #1198
• Preserve Drama3 history anchors and fit generation to context by @0xShug0 in #1197
• Add D1-Omni audio decisions and reusable LFM runtimes by @0xShug0 in #1199
• Add opt-in skipping of empty Drama3 chunks by @0xShug0 in #1200
• Refresh model usage guides and consolidate validation reports by @0xShug0 in #1201
• Prepare WebUI 0.9.4 with missing models and app readiness by @0xShug0 in #1202
• Add WebGPU support for 3D im2col by @0xShug0 in #1203
• v0.9.4 cleanup by @0xShug0 in #1204

New Contributors

• @linyqh made their first contribution in #1108
• @luizribeiro made their first contribution in #1125
• @supomarlo made their first contribution in #1146
• @askmyteapot made their first contribution in #1143
• @ruihe774 made their first contribution in #1152
• @m-abs made their first contribution in #1148

Full Changelog: v0.9.3...v0.9.4