VoiceChat configuration
The realtime server loads one process-level VoiceChat runtime from a converted model-version directory. s2s.model_dir enables the runtime automatically; s2s.enabled can override automatic detection.
Common settings
KeyDefaultDescriptions2s.enabledautoEnablement policy: auto, true, or falses2s.model_diremptyConverted model-version directorys2s.max_streams32Maximum number of resident conversation statess2s.verbosefalseEmit detailed GGML and llama.cpp diagnosticsnemo-speech serve also accepts --s2s-model-dir and --s2s-max-streams as aliases.
s2s: enabled: automodel_dir: /models/NVIDIA-NemotronLabs-VoiceChat-11B-GGUFmax_streams: 32verbose: falsenemo-speech serve --config config/voicechat.yamls2s.max_streams is a state-reservation ceiling, not a throughput target. Set it to 1 for a single-conversation deployment. Incoming requests are batched dynamically; conversation, sampler, and generated-audio state remain isolated per stream.
When nemo-speech serve chooses its default HTTP worker count, it reserves enough workers for the configured VoiceChat stream ceiling. An explicitly set http.threads value remains authoritative; it should be at least the intended number of simultaneous WebSocket sessions.
Realtime WebSocket settings
These settings apply to nemo-speech serve:
KeyDefaultDescriptions2s.max_session_seconds300Maximum cumulative input-audio duration per sessions2s.max_pending_function_responses64Maximum queued tool responses per sessions2s.output_text_eventsfalseAlso emit legacy response.output_text.* eventsThe listener's http.api-key, upload limit, read timeout, write timeout, and TLS settings apply to VoiceChat sockets. Never put the configured, long-lived API key in a URL or the api_key query parameter, where logs and browser history can expose it. Header-capable clients should send Authorization: Bearer <key>. Browser clients should obtain a short-lived, scope-limited token from a trusted backend instead of receiving the configured server key.
Runtime environment variables
VariableDefaultDescriptionS2S_BATCH_QUEUE_DELAY_US1000Maximum internal stage batching delayS2S_INGRESS_COHORT_DELAY_US2000Initial stream-cohort collection delayS2S_TTS_SEED0Acoustic-token sampling seed; nonzero values enable reproducible samplingS2S_MAX_STREAMSconfigured valueOverride the resident-stream ceilingS2S_DEBUG_TIMINGunsetEmit per-stage timing diagnosticsEnvironment variables are optional. Start with defaults and change batching delays only when tuning for a measured workload.