v0.9.0

What's Changed

• Speed up BS-RoFormer CUDA separation by batching projections by @mirek190 in #666
• higgs_audio_tts: optimize CUDA KV-cache transitions by @mirek190 in #669
• pocket_tts: use flow_lm.flow.depth for the flow net (fixes *_24l loading) by @bvisible in #670
• pocket_tts: chunk text like the reference, UTF-8 aware text prep by @bvisible in #671
• moss_tts_v15: call the backbone's own timing report by @christopherthompson81 in #676
• Improve Nemotron 3 diarization NeMo parity by @LysanderdeJong in #677
• Add Pascal support to Windows CUDA 12.4 releases by @0xShug0 in #682
• moss_ttsd: MOSS-TTSD, and share the moss_tts_delay prompt assembly by @christopherthompson81 in #652
• Reduce MOSS-TTS v1.5 GPU memory usage by @0xShug0 in #674
• voxtral_realtime: never end a streaming delta inside a UTF-8 character by @christopherthompson81 in #684
• moss_ttsd: take #674's f16 decoder and KV cache, keep the encoder at f32 by @christopherthompson81 in #687
• moss_ttsd: hold the encoder at its stored type and widen to f32 in the graph by @christopherthompson81 in #690
• metrics: process peak memory in --metrics, reserved-vs-used per ggml context in --log by @engival in #688
• nemotron_asr: move options to the model spec by @LysanderdeJong in #685
• fix(webui): client side clone voice transcript handling too strict by @jasonchen31 in #693
• feat(zipvoice): Add voice streaming to ZipVoice by @jasonchen31 in #692
• docs: add Remiqora to the Projects list by @inikolax in #695
• nemotron_3_diar: write speaker probabilities as safetensors by @LysanderdeJong in #686
• confucius4_r2t2: make streaming chunks incremental instead of O(utterance) by @scriptease in #696
• Add speaker-tagged transcription to Nemotron ASR by @LysanderdeJong in #681
• confucius4_r2t2: add community Q4_K_M GGUF package (pinned, no re-hosting) by @NairoDorian in #702
• Keep published Nemotron diarization GGUFs compatible by @0xShug0 in #703
• Higgs audio v3: reduce CUDA prefill peak VRAM 91% with graph allocation! by @mirek190 in #704
• higgs_audio_tts: reuse prefill storage on Vulkan by @0xShug0 in #705
• Add HTDemucs 6-stem source separation (drums, bass, vocals, other, guitar, piano) by @SMGoro in #694
• Add optimized Maya1 TTS support by @0xShug0 in #707
• Add GigaAM v3 and multilingual ASR support by @0xShug0 in #708
• Add native SAMSONE audio understanding by @0xShug0 in #709
• Add native SAM Audio separation by @0xShug0 in #711
• confucius4_r2t2: fix null deref when the streaming final flush has no tokens by @hdcola in #712
• Add standalone Tone Color VC model support by @0xShug0 in #716
• Add Japanese translation for WebUI by @asfdrwe in #719
• Enable Higgs Audio TTS prefill storage reuse on HIP by @0xShug0 in #722
• Preserve CORS on errors and document Qwen3 ASR timestamps by @0xShug0 in #721
• Rename shared HuBERT encoder as Wav2Vec2 runtime by @0xShug0 in #724
• Clarify shared decoder and model architecture names by @0xShug0 in #725
• Make model-local component architecture names explicit (1) by @0xShug0 in #727
• Make model-local component architecture names explicit (2) by @0xShug0 in #728
• Make model-local component architecture names explicit (3) by @0xShug0 in #729
• Echo-TTS VRAM improvements by @lojack5 in #731
• Organize framework components by responsibility by @0xShug0 in #734
• echo_tts: #731 follow up - compatibility with older embeded model specs by @lojack5 in #736
• ACE-Step family : reduce memory usage 63% and speedup for vulkan 16% and cuda 6% by @mirek190 in #737
• Add new models and enhancement tools to the WebUI by @0xShug0 in #738

New Contributors

• @bvisible made their first contribution in #670
• @inikolax made their first contribution in #695
• @NairoDorian made their first contribution in #702
• @SMGoro made their first contribution in #694
• @hdcola made their first contribution in #712
• @asfdrwe made their first contribution in #719
• @lojack5 made their first contribution in #731

Full Changelog: v0.8.2...v0.9.0