Pull requests

• fix(paraformer): keep one keyed result per utterance when no tokens are predicted

  #3797·MohammadHijjawi97 opened Oct 27, 2026Contributor·2 of 9 tasks·

  ·
  1

• fix(diarization): honor configurable speaker caps for long recordings

  #3794·phoenixray2000 opened Oct 24, 2026Contributor·

  ·

• fix(auto): preserve Paraformer token alignment across VAD segments

  #3793·phoenixray2000 opened Oct 24, 2026Contributor·7 of 12 tasks·

  ·

• fix(onnx): return an empty list for empty split-final Paraformer results

  #3792·ychy7001 opened Oct 24, 2026·2 of 6 tasks·

  ·
  1

• fix(onnx): keep no left-context frames when chunk_size[0] is 0 in streaming Paraformer

  #3790·ychy7001 opened Oct 24, 2026·4 of 9 tasks·

  ·
  1

• fix(auto): preserve speaker dependency import failures

  #3788·monsterlag opened Oct 22, 2026·2 of 9 tasks·

  ·

• fix(train): select XCCL as the XPU distributed backend instead of NCCL

  #3786·li-lizhe opened Oct 21, 2026Contributor·

  ·
  1

• fix(tokenizer): respect the model vocabulary in ids2tokens

  #3783·carey-bk opened Oct 20, 2026·

  ·

• Prepare FunASR 1.4.17 release candidate

  #3770·LauraGPT opened Oct 10, 2026Collaborator·

  ·

• feat: support hf-overrides for realtime server

  #3768·gotorion opened Oct 9, 2026Contributor·4 of 9 tasks·

  ·

• fix(fun-asr-nano): restore repetitive-garbage trimming in streaming vLLM

  #3766·MincongZhou opened Oct 8, 2026·2 of 9 tasks·

  ·

• fix(cli): warn when jieba is unavailable for readable SRT splitting

  #3726·Linxiushen opened Sep 21, 2026·

  ·
  1

• fix(realtime): decouple bounded receive buffering from inference

  #3686·LauraGPT opened Sep 9, 2026Collaborator·

  ·