v0.7.2

What's Changed

• docs(server): explain host and port options by @mirek190 in #381
• fix(granite5asr): build mel filterbank on the continuous frequency axis by @gsaon in #384
• feat: Supporting Audio8_TTS models by @jasonchen31 in #333
• moss_voicegen: accept the instruction spelling the speech route sends by @CryptVenture in #371
• Fix ACE-Step auto-duration planner budget by @0xShug0 in #390
• audio: add 24-bit and float32 WAV output, opt-in dither, and a limiter by @CryptVenture in #358
• model_spec: correct the shared text_chunk_mode preset by @CryptVenture in #370
• Add multipart audio alignment endpoint by @0xShug0 in #391
• feat: support set mirror by env:HF_ENDPOINT when downloading models by @feng19 in #397
• perf(fish_audio): add HIP Fast-AR top-k sampler by @reezex0-ux in #386
• server: forward language only where the model's contract accepts it by @CryptVenture in #400
• fix(voxcpm1): fix webui download failure and Yue lang mis-triggering issues by @jasonchen31 in #424
• fix(qwen): preserve compact logits indices across graph reuse by @mirek190 in #426
• Add opt-in Qwen3 ASR punctuated text output by @0xShug0 in #433
• feat: add Chatterbox Turbo TTS model family by @pannagaps in #394
• Fix Higgs Audio codec decode seams by @0xShug0 in #436
• Reduce Vevo2 FM graph peak memory by @0xShug0 in #443
• fix(text-chunking): split Default-mode CJK text at full-width punctuation by @gqf2008 in #441
• Fix soprano_tts maintainer information in README by @drzsdrtfg in #451
• Merge Breeze and CosyVoice3 from dev into main by @0xShug0 in #452
• Fix CosyVoice3 Metal flow layout by @0xShug0 in #455

New Contributors

• @gsaon made their first contribution in #384
• @feng19 made their first contribution in #397
• @reezex0-ux made their first contribution in #386
• @pannagaps made their first contribution in #394

Full Changelog: v0.7.1...v0.7.2