| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677 |
- [audio]
- backend = "parec" # "parec" (PulseAudio by name) | "sounddevice"
- device = "VirtualMicSink.monitor" # PulseAudio source name (parec) or device index/name (sounddevice)
- sample_rate = 16000
- channels = 1
- frame_ms = 30 # webrtcvad requires 10/20/30 ms
- gain = 1.0
- [vad]
- aggressiveness = 2 # webrtcvad 0-3
- pre_roll_ms = 400 # ring buffer kept BEFORE speech is confirmed
- speech_start_ms = 90 # consecutive speech frames needed to confirm start
- silence_ms = 700 # trailing silence that ends a turn
- min_utterance_ms = 300 # discard utterances with less actual speech than this
- max_utterance_s = 30.0 # hard cap: force-finalize a very long utterance
- chunk_pause_ms = 350 # once soft_max_s is reached, end a chunk at a pause this long
- soft_max_s = 10.0 # after this much speech, chunk-ends become eligible at pauses
- rms_floor = 0.008 # drop segments quieter than this RMS (kills noise-triggered blips)
- trailing_trim_ms = 200 # keep only this much trailing silence (less hallucination-prone padding)
- [whisper]
- model = "small" # multilingual (DO NOT use .en suffix)
- device = "cpu" # "cpu" | "cuda"
- compute_type = "int8" # "int8" cpu, "float16" cuda
- language = "auto" # "auto" | "en" | "es"
- beam_size = 1
- vad_filter = false
- no_speech_prob_max = 0.6 # drop segments Whisper thinks are non-speech above this
- avg_logprob_min = -1.0 # drop segments with confidence below this
- [output]
- stdout = true
- jsonl_dir = "transcripts" # one JSONL file per day: YYYY-MM-DD.jsonl
- save_audio = false # write each transcribed utterance as a 16 kHz mono WAV
- audio_dir = "audio" # WAV destination (YYYY-MM-DD subfolders); used only if save_audio
- [sink]
- sink_name = "VirtualMicSink" # Name of the null sink to create
- remap_source_name = "VirtualMic" # Name of the exposed microphone
- loopback_latency_msec = 1 # Loopback latency in ms
- monitor_output = false # Loop sink.monitor to default sink (AVOID: causes acoustic feedback)
- [daemon]
- queue_size = 50 # max utterances waiting for transcription (oldest dropped when full)
- backlog_warn_at = 5 # warn when the queue backs up this far (worker falling behind)
- [events]
- socket = "" # publish-only event socket (default $XDG_RUNTIME_DIR/stt-daemon.sock)
- heartbeat_ms = 1000 # heartbeat interval (progress-driven, from the capture loop)
- ring_size = 200 # recent events replayed to a newly attached subscriber
- [assistant]
- control_socket = "" # control socket (default $XDG_RUNTIME_DIR/stt-assistant.sock)
- stall_timeout_ms = 3000 # no heartbeat for this long -> STT considered offline
- worker_stall_ms = 15000 # heartbeats arriving but a transcription runs this long -> degraded
- command_arm_timeout_s = 10 # disarm only if armed with NOTHING captured for this long
- chat_idle_timeout_s = 300 # leave chat mode after this much silence
- turn_gap_s = 1.5 # quiet gap that ends a turn
- require_viewer = true # refuse to execute commands with no TUI attached
- chat_exit_phrases = ["end conversation", "that's all", "stop listening"]
- [audit]
- enabled = true # write executed actions to a daily JSONL file
- dir = "audit" # audit/YYYY-MM-DD.jsonl
- [llm]
- enabled = true
- endpoint = "http://localhost:9931/v1/chat/completions" # llama.cpp, OpenAI-compatible
- temperature = 0.7
- max_tokens = 1024 # reasoning models spend tokens before the answer; keep this generous
- timeout_s = 120
- cache_prompt = true # reuse the KV prefix (append-only history keeps it warm)
- max_prompt_tokens = 64000 # ~half of the 131072 context
- trim_block_turns = 4 # drop history in blocks (not per-turn) to limit re-prefill
- max_tool_iterations = 3
- system_prompt_chat = "You are a concise voice assistant. Reply in 1-3 sentences."
- system_prompt_command = "You translate a spoken request into a single tool call when one applies."
|
