Skip to content

v0.9.6

Latest

Choose a tag to compare

@github-actions github-actions released this 21 Nov 18:12
· 4 commits to main since this release

What's Changed

  • README: add Woogles-TV to Projects by @jgbrwn in #1286
  • Fix PocketTTS streaming quality and speed up PocketTTS decoders by @0xShug0 in #1291
  • cmake: add AUDIOCPP_BUILD_C_API_STATIC option for embedding (libaudiocpp.a) by @codixor in #1290
  • campplus: write the graph's constants before every run, not only the first by @magnus-dev-melin in #1284
  • chatterbox: keep the component and its voices across languages that share a T3 by @magnus-dev-melin in #1285
  • Expose custom-schema classification results through the C API by @0xShug0 in #1292
  • Share MoE FFN and Gated DeltaNet graph construction by @0xShug0 in #1293
  • Add M5-TTS voice cloning, preset synthesis and streaming by @0xShug0 in #1298
  • omnivoice: run the generator on the vendor zero-op matmul API (rknn_zeroop) by @wOvAN in #1297
  • omnivoice: harden the rknn_zeroop packed-weight disk cache by @wOvAN in #1302
  • omnivoice: review-round fixes for the rknn_zeroop weight cache by @wOvAN in #1303
  • Fix false Xet checksum mismatches on Windows by @0xShug0 in #1301
  • Add ACE-Step reference-audio conditioning by @0xShug0 in #1307
  • Add ROSVOT singing-note transcription by @0xShug0 in #1308
  • Refactor Conformer blocks to reuse shared feed-forward modules by @0xShug0 in #1309
  • feat: add tiny-voca, a 4.5M-parameter English TTS model with its own phonemizer by @n4ze3m in #1311
  • Add YAML config support to CLI by @j0n4t in #1304
  • feat: detect tensor container by content when there is no extension by @kikaitachi in #1315
  • Add Linux and Windows HIP rocBLAS compatibility packages by @0xShug0 in #1316
  • Fix CLI precedence for MoE YAML options by @j0n4t in #1317
  • Fix PocketTTS decoder lifetimes across chunk and session reuse by @0xShug0 in #1314
  • Consolidate PocketTTS Small Mimi decoding and reduce inference overhead by @0xShug0 in #1318
  • Add SoulX-Singer singing synthesis and voice conversion by @0xShug0 in #1319
  • 🚨 Use shared Mimi decoding for PocketTTS and correct streaming history by @0xShug0 in #1320
  • Fix RMVPE stream lookup with dynamic CUDA backends by @0xShug0 in #1321
  • fix(tiny_voca): keep graphs in a bounded session cache and refuse languages other than English by @n4ze3m in #1324
  • Preserve Japanese text during Chinese-script normalization by @0xShug0 in #1326
  • Continue ZONOS2 long-form chunks from earlier audio codes for voice consistency by @Ryu1845 in #1328
  • Add Latin support by @coolcoder613eb in #1322
  • Reduce CosyVoice3 memory reservations and add speech speed control by @0xShug0 in #1330
  • Bound HiFT graph metadata memory independently of audio length by @0xShug0 in #1331
  • 🚨 Introduce a standalone MoE engine with independent placement and cache policies by @0xShug0 in #1334
  • Refactor Granite5ASR graph reuse and batch GPU attention by @0xShug0 in #1335
  • New Community Model: Granite Speech 5.0 2B (IBM) by @eutychus in #1332
  • Fix MuLaCover ChordNet allocation for long references by @0xShug0 in #1344
  • pocket_tts: apply the config's replace_characters like the reference by @unaltolga in #1339
  • fish_drama3: keep several encoded reference voices (LRU) by @grogroningem-code in #1345
  • Reduce oversized GGML metadata arena reservations (1) by @0xShug0 in #1346
  • Reduce oversized GGML metadata arena reservations (2) by @0xShug0 in #1347
  • Fix Qwen3 Japanese/Korean alignment and add opt-in overlap text trimming by @0xShug0 in #1350
  • pocket_tts: give the built-in defaults the upstream text options by @unaltolga in #1352
  • Add VoxMel TTS and voice cloning by @0xShug0 in #1353
  • Add missing models to WebUI and bump UI to 0.9.6 [skip ci] by @0xShug0 in #1354
  • perf(kokoro): reduce CPU inference time and graph memory by @0xShug0 in #1355
  • Refactor Qwen3.5 runtime to reuse shared modules by @0xShug0 in #1356
  • ggml-cpu: speed up Mamba-1 and quantized-V flash attention; add an opt-in GELU erf approximation by @Etherll in #1342
  • New Community Model: Voice Isolator v1 (Inworld) by @Etherll in #1343
  • Improve Voice Isolator performance, streaming reuse and WebUI/WebGPU support by @0xShug0 in #1358
  • Fix Voice Isolator offline Metal projection overflow by @0xShug0 in #1359
  • Extend WebGPU operation support and browser model validation by @0xShug0 in #1361
  • Add opt-in Kokoro phoneme rewrite rules by @0xShug0 in #1360

New Contributors

Full Changelog: v0.9.5...v0.9.6