v0.8.2

What's Changed

• tests(yue2): add a standalone NAR parity probe by @mirek190 in #598
• Add explicit ggml lowering APIs by @0xShug0 in #532
• Add CUDA lowering implementations by @0xShug0 in #533
• Add opt-in CUDA SSM gate fusion and tiled F32 im2col by @0xShug0 in #600
• Add reusable Wan S2V framework components by @0xShug0 in #601
• Add LiveAvatar native model support by @0xShug0 in #605
• fix(scripts): guard empty arrays so the build scripts survive bash 3.2 by @gqf2008 in #602
• breeze: enable bf16 activation rounding on Metal by @gqf2008 in #554
• zipvoice: add k2-fsa ZipVoice community TTS with zh/en frontend by @davidxifeng in #599
• feat(model): add Confucius4-R2T2 real-time streaming ASR (r2t2_asr) by @davidxifeng in #604
• vibevoice_asr_streaming: add the 1.5B size, which the loader already handles by @christopherthompson81 in #613
• Add AuK Base and AuK-Flash community model by @0xShug0 in #615
• Support speech speed in server and TTS model interfaces by @0xShug0 in #616
• Rebuild an ASR encoder graph when the cached one is too big by @christopherthompson81 in #619
• fix(voxcpm2): decode streaming patches with left context by @mduka in #622
• Accept a prompt query parameter on the live transcription route by @scriptease in #618
• fix(voxcpm2): correct the streaming decode rationale, drop right context by @mduka in #623
• Add Piper and KittenTTS community models by @0xShug0 in #625
• fix(voxcpm1): use the stateful streaming decode on Vulkan too by @mduka in #624
• vulkan-shaders-gen: retry an empty shader compile instead of failing at link (potential fix for CI failures) by @christopherthompson81 in #540
• webui: make the committed bundle reproducible by @christopherthompson81 in #546
• ci: run the CI jobs locally, and actually run the unit tests on linux by @christopherthompson81 in #627
• Share a configurable NeMo mel frontend across ASR models by @0xShug0 in #628
• moss_tts_v15: MOSS-TTS-v1.5, and promote the moss_tts_delay family runtimes by @christopherthompson81 in #631
• Improve Nemotron ASR streaming by @0xShug0 in #630
• Share mel and reference audio frontends by @0xShug0 in #629
• kokoro_tts: report the durations it already predicts as word timings by @christopherthompson81 in #626
• yue2: export the semantic token stream as a result artifact by @engival in #635
• perf(roformer): pipeline CUDA chunk processing by @mirek190 in #633
• webui: wire the YuE2 NAR LoRA adapter into the UI by @christopherthompson81 in #614
• Yue2 "stop after" (generate just score.abc, or semantic) by @engival in #636
• yue2: tile the eager NAR attention so long songs fit a 4 GiB buffer [bugfix] by @engival in #642
• vulkan-shaders-gen: retry a compiler that cannot be started, and back off under memory pressure by @christopherthompson81 in #634
• feat(audio8_tts): Falcon-H1 0.1B (Mamba2 + attention) support and codec/AR performance by @gqf2008 in #444
• perf(cuda): optimize narrow RoFormer operations by @mirek190 in #632
• docs: add a license overview for model weights by @jrohde in #648
• vieneu_v3_turbo (formerly vietneu_tts): reference-code conditioning, Python sampling parity, official GGUF — from the model author by @pnnbao97 in #647
• yue2: continue a song from given semantic tokens by @engival in #641
• gguf: quantize a tensor's rows in parallel by @christopherthompson81 in #649
• moss_tts_v15: declare the text chunking options in the spec by @christopherthompson81 in #654
• Fix Seed-VC V1 F0 conditioning by @0xShug0 in #653
• fix(workflow): escape control characters in file-sink JSON by @DivyamTalwar in #640
• Update WebUI for v0.8.2 by @0xShug0 in #655
• vieneu_v3_turbo: chunking and pauses, Vietnamese text input, enrolment artifacts, f16 codec in both packages by @pnnbao97 in #658
• fix(moonshine_asr): reject audio beyond the position-embedding limit instead of crashing by @Faazil-jpg in #662
• readme: add AudioCpp-Bindings to Projects by @christopherthompson81 in #665
• moss codec: the Nano decoder attended over twice its intended window (fixes #663) by @christopherthompson81 in #664
• Add Nemotron 3 diarization support by @0xShug0 in #660

New Contributors

• @davidxifeng made their first contribution in #599
• @mduka made their first contribution in #622
• @scriptease made their first contribution in #618
• @pnnbao97 made their first contribution in #647
• @DivyamTalwar made their first contribution in #640
• @Faazil-jpg made their first contribution in #662

Full Changelog: v0.8.1...v0.8.2