# nemo-speech ASR config. Load with: nemo-speech serve --config this.yaml # # Nested maps mirror dotted command-line keys: `asr.vad.masker.onset` is the # same setting as `--asr.vad.masker.onset`. See docs/asr/configuration.md for # the complete key reference. # # Precedence is: this file < NEMO_SPEECH_* env vars < CLI flags. # Every key is optional - omitted keys keep their built-in defaults. An UNKNOWN # key is a hard error (so typos are caught, not silently ignored). The values # shown here are the defaults unless noted; the model reference and commented # paths are examples. asr: backend: gpu: 0 # GPU device index, -1 = CPU model: path: nemotron-3.5 # indexed name, HF repo ID, or local GGUF # name: # display name; default is derived from the model streaming: chunk_size: 0.16 # CTC buffered window (seconds) ctc_left_padding: 1.92 # CTC left context (seconds) ctc_right_padding: 1.92 # CTC right context (seconds) rnnt_right_context: 1 # cache-aware R: 1 = low-latency preset, -1 = model max batching: enabled: true # nemo-speech serve default; disable to minimize single-request latency max_batch_size: 1024 max_queue_delay_us: 5000 state_arena_slots: 16 offline_bucket_ms: 0 # 0 = do not silence-pad offline inputs into shared buckets decoder: kind: greedy # greedy | flashlight # Flashlight LM beam (set kind: flashlight, or just lm_path which implies it): # lm_path: /models/lm.bin # lexicon_path: /models/lexicon.txt # tokenizer_path: /models/tokenizer.model # for OOV word boosting # beam_size: 32 # beam_size_token: 16 # beam_threshold: 20.0 # lm_weight: 0.8 # word_insertion_score: 1.0 vad: # model_path: /models/silero-v6.2.0.gguf # empty = no VAD loaded masker: mask_enable: false # mask silence features before the encoder onset: 0.5 # prob > onset -> enter speech offset: 0.3 # prob < offset -> leave speech # pad_onset_ms: 200 # pad_offset_ms: 200 # min_duration_off_ms: 500 endpointing: enable: false # mid-stream end-of-utterance vad_based: false # ride the VAD timeline (needs vad.model_path); else token-silence stop_history_eou_ms: 800 # trailing silence before EOU (ms) diar: # model_path: sortformer # enables word-level speaker tags when requested preset: streaming # streaming | offline (larger chunks and caches) postproc: # pnc_model_path: /models/pnc.gguf # punctuation + capitalization # itn_model_dir: /models/sparrowhawk_en # inverse text normalization # profanity_list_path: /models/profanity.txt