# nemo-speech combined HTTP server config. # # Load with: # nemo-speech serve --config config/server.example.yaml # # ASR, diarization, NMT, and TTS are enabled when their required model paths # are present. Set `enabled: false` under ASR, NMT, or TTS to force it off. http: enabled: true host: 127.0.0.1 port: 8080 threads: 4 max-upload-mb: 512 read-timeout: 30 write-timeout: 30 access-log: false log-format: text # text | json playground: true # api-key: use NEMO_SPEECH_HTTP_API_KEY instead of storing secrets here # cors-origin: https://trusted-client.example # tls-cert: /run/secrets/server.crt # tls-key: /run/secrets/server.key asr: enabled: true # true | false | auto backend: gpu: 0 # GPU device index, -1 = CPU model: path: /models/nemotron-speech-streaming-en-0.6b.q8_0.gguf streaming: chunk_size: 0.16 ctc_left_padding: 1.92 ctc_right_padding: 1.92 rnnt_right_context: 1 decoder: kind: greedy # greedy | flashlight vad: model_path: # empty = no VAD loaded masker: mask_enable: false onset: 0.5 offset: 0.3 endpointing: enable: true vad_based: false stop_history_eou_ms: 800 postproc: # pnc_model_path: /models/pnc.gguf # itn_model_dir: /models/sparrowhawk_en # profanity_list_path: /models/profanity.txt # This model also enables word-level speaker tags on ASR responses. diar: model_path: nvidia/diar_streaming_sortformer_4spk-v2 preset: streaming # streaming | offline (larger chunks and caches) tts: enabled: true # true | false | auto magpie-model: /models/magpie-tts/magpie_tts_multilingual_357m.v2602.f16.gguf codec-model: /models/nano-codec/nemo_nano_codec_22khz_1.89kbps_21.5fps.decoder.f16.gguf tokenizer-model-dir: /models/magpie-tts/extracted # tn-model-dir: /models/en_tn_grammars_cased # written-form -> spoken-form TN language-code: en-US speaker: 0 threads: 4 codec-threads: 0 # 0 = use threads seed: -1 # -1 = current time steps: -1 # -1 = model default top-k: -1 # -1 = model default chunk-frames: 4 codec-queue-depth: 4 codec-history-frames: -1 codec-future-frames: 1 window-ms: 0 flush-partial-chunk: true use-cfg: true use-local-transformer: true use-kv-cache: true use-stateful-codec: true codec-cpu: false lt-backend: auto # auto | cpu | cuda sampling-backend: auto # CUDA with a CUDA Magpie/LT path; otherwise CPU uma-mode: auto # auto | off | on longform: auto # auto | off | on voice-name: warmup-enabled: true warmup-text: "Hello from Magpie T T S." warmup-steps: 8 # NMT (text translation) is built only when -DNEMO_SPEECH_BUILD_NMT=ON. Like ASR # and TTS it is enabled automatically when its model path is present. nmt: enabled: true # true | false | auto backend: gpu: 0 # GPU device index, -1 = CPU model: path: /models/riva-translate-4b-instruct-v2.q8_0.gguf # f16 also works n_ctx: 1024 # raise for longer inputs; KV cache grows with n_ctx generation: max_new_tokens: 256 pool: contexts: 1 # concurrent decode contexts