Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
docs: clarify llama.cpp speaker diarization support
zhifu gao65bbd29f603c
diff --git a/runtime/llama.cpp/README.md b/runtime/llama.cpp/README.mdindex 5d5ee2c9a..ddfe1f210 100644--- a/runtime/llama.cpp/README.md+++ b/runtime/llama.cpp/README.md@@ -78,6 +78,19 @@ segmentation, SenseVoiceSmall and Paraformer emit one entry spanning the input.Progress and timing diagnostics remain on stderr, so redirecting stdout producesa clean subtitle file. Normal text output is unchanged when `--srt` is omitted.+## Speaker diarization++The standalone llama.cpp / GGUF binaries do **not** currently implement CAM+++speaker embeddings or speaker clustering. `--vad` segments speech for long-audio+transcription, but it does not assign speaker labels.++For speaker-aware transcripts, use the Python `AutoModel` pipeline with+`spk_model="cam++"` (see the [FunASR quick start](../../README.md#quick-start)) or+the [Fun-ASR-Nano vLLM service](../../examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py),+which accepts `spk=true` and runs a separate speaker model. Keep using the+llama.cpp binaries when the requirement is self-contained ASR plus optional+FSMN-VAD without Python.+### Optional Windows CUDA backend for SenseVoiceSmallThe CPU release ZIPs are portable packages. Tagged releases also publish