v0.9.1

Recent release builds tend to target AVX-512 due to automatic CPU detection on GitHub runners. If you run into crashes, please use the portable version for broader CPU compatibility.

What's Changed

• Add opt-in SSM and projection support for CUDA and Vulkan by @0xShug0 in #713
• Add RE-USE audio restoration and native task batching by @0xShug0 in #714
• fix(reuse): disable HIP and Metal backends by @0xShug0 in #741
• Add Sidon speech restoration model by @0xShug0 in #742
• Add Smart Turn v3.2 offline turn detection by @0xShug0 in #743
• Add KugelAudio native TTS with batched CFG and streaming by @0xShug0 in #744
• fix(server): support inline Base64 audio in task requests by @0xShug0 in #750
• Restore executable bit on release tarball binaries by @scriptease in #754
• fix(model-manager): reduce inventory DNS lookups by @0xShug0 in #752
• Add Index-Echo S2TT translation models by @0xShug0 in #755
• fix(safetensors): bound nesting depth when skipping metadata by @Zhuoxi2000 in #753
• fix(torch_bin): honour _rebuild_tensor_v2 storage_offset and reject non-contiguous tensors by @Zhuoxi2000 in #757
• feat: add lfm2_audio ASR for LFM2.5-Audio EN and JP by @ykhrustalev in #758
• Add CrisperWhisper 2.0 with long-form transcription and alignment by @0xShug0 in #759
• Add Audio Flamingo 3 and Next support by @0xShug0 in #763
• Lower YuE2 context defaults by @0xShug0 in #767
• fix(lfm2_audio): end a transcript cut at max_tokens at the last whole character by @ykhrustalev in #765
• perf(lfm2_audio): keep the encoder graph and its compute buffer between chunks by @ykhrustalev in #766
• docs(r2t2): note that the final word needs trailing silence by @scriptease in #770
• Use UTF-8 process code pages for Windows executables by @0xShug0 in #772
• Add OWSM v4 and OWSM-CTC v4 by @0xShug0 in #774
• server: separate legacy and experimental parallel jobs runtimes by @mirek190 in #715
• Fix Chatterbox S3 encoder relative shift and LayerNorm epsilon by @dignome in #778
• Add Kittentts2 community model by @dignome in #776
• Fix WebUI Run for bundled VAD models by @mister2d in #782
• Fix Kitten TTS 2 reported wall_ms and clean up model implementation by @dignome in #781
• Egyptian arabic voicetut qwencleo by @MohammedAly22 in #780
• feat: add lfm2_audio TTS for LFM2.5-Audio EN and JP by @ykhrustalev in #768
• Align UI catalog model paths with spec packages by @0xShug0 in #783
• Add tts-audiobook-tool to projects listing in README.md by @zeropointnine in #788
• Reduce HTDemucs inference overhead across backends by @0xShug0 in #790
• Document OmniVoice and Qwen3 options with compatible normalized names by @0xShug0 in #786
• server: add include_params to GET /v1/models, read from the model spec by @Navier8 in #787
• Fix repeated Voxtral Realtime transcriptions by @dignome in #793
• Enhance CUDA 12 library handling in Colab notebook by @ha466 in #795
• fix: allow --out to target FIFOs and char devices (e.g. /dev/stdout) by @trappedinspacetime in #796
• feat: add lfm2_audio speech-to-speech for LFM2.5-Audio EN and JP by @ykhrustalev in #798
• docs: add ComeCut to the Projects list by @juntaosun in #804
• Use depthwise upsampling for PocketTTS on CPU by @0xShug0 in #806
• feat: return_text on /v1/audio/speech/live by @ykhrustalev in #805
• 🚨 Accept neutral speed in OpenAI speech requests by @0xShug0 in #807
• Add BreezeTTS2 Q4 package and PocketTTS voice export docs by @0xShug0 in #808
• Fix ZipEnhancer corrupted enhancement output by @0xShug0 in #811
• Document core model options with compatible normalized names by @0xShug0 in #814
• Fix Stable Audio short-input encoder stitching by @0xShug0 in #815
• refactor(lfm2_audio): choose the decode-cache size by task by @ykhrustalev in #813
• docs(reports): add YuE2 longform host commit exhaustion report by @wangguanghua2099 in #816
• Sopro v2.2.0 parity, real (chunked) streaming, and opt-in voice cache by @samuel-vitorino in #818
• Relax Niagara ASR and AuK backend restrictions by @0xShug0 in #817
• fix(lfm2_audio): refuse audio too long for one encoder pass, and name missing package files by @ykhrustalev in #820
• Reduce Qwen3-TTS, IndexTTS, and BreezeTTS memory allocations by @0xShug0 in #821
• server: optional input on /v1/audio/speech/live for s2s and vc models by @ykhrustalev in #822
• feat(lfm2_audio): text_temperature and text_top_k for S2S by @ykhrustalev in #827
• v0.9.1 WebUI Model Pass by @0xShug0 in #819
• fix(webui): no text box for lfm2_audio speech-to-speech by @ykhrustalev in #828

New Contributors

• @Zhuoxi2000 made their first contribution in #753
• @ykhrustalev made their first contribution in #758
• @dignome made their first contribution in #778
• @mister2d made their first contribution in #782
• @MohammedAly22 made their first contribution in #780
• @zeropointnine made their first contribution in #788
• @Navier8 made their first contribution in #787
• @ha466 made their first contribution in #795
• @trappedinspacetime made their first contribution in #796
• @juntaosun made their first contribution in #804
• @wangguanghua2099 made their first contribution in #816
• @samuel-vitorino made their first contribution in #818

Full Changelog: v0.9.0...v0.9.1