Skip to main content
Continuous voice mode listens for speech, transcribes, processes commands, and speaks responses — all without manual intervention. Ideal for hands-free workflows.

​
Usage

rcli listen

# With custom options
rcli listen --no-speak --gpu-layers 50

# With RAG index loaded
rcli listen --rag ~/Library/RCLI/index

​
How It Works

  1. Always Listening — VAD (Silero) detects speech activity
  2. Streaming STT — Zipformer transcribes live audio
  3. Endpoint Detection — Detects silence, finalizes transcript
  4. LLM Processing — Classifies intent, executes actions, generates response
  5. TTS Playback — Speaks the result via Piper/Kokoro
  6. Loop — Returns to listening state

​
Voice Pipeline States

The terminal displays an animated dog indicating the current state: 
  /^ ^\
 ( o.o )
  > ^ <
   |_|

Listening for speech...

​
Supported Commands

​
macOS Actions

"open Safari"
"set volume to 50"
"create a note called Meeting Notes"
"play some jazz on Spotify"
"take a screenshot"
"lock screen"

​
Conversational Queries

"what is the weather today?"
"explain quantum computing"
"what's on my calendar?"

​
RAG Queries (with —rag)

"what were the key decisions from the meeting?"
"summarize the project plan"

​
Options

​
--models
string
default:"~/Library/RCLI/models"
Models directory path
​
--gpu-layers
number
default:"99"
GPU layers for LLM (99 = all, 0 = CPU only)
​
--no-speak
boolean
Disable TTS audio output (text only)
​
--rag
string
Load RAG index for document-grounded answers
​
--verbose
boolean
Show debug logs from engines

​
Voice Activity Detection (VAD)

Listen mode uses Silero VAD to:
  • Filter background noise
  • Detect speech start/end
  • Trigger transcription only when speaking
This reduces false positives and improves transcription accuracy.

​
STT Models

Listen mode uses two STT models in parallel:

​
Zipformer (Streaming)

  • Purpose — Real-time transcription during speech
  • Speed — ~50ms latency
  • Accuracy — Good for live feedback
  • Size — ~50 MB

​
Whisper/Parakeet (Offline)

  • Purpose — Final accurate transcription after speech ends
  • Speed — ~40ms for Whisper base.en, ~60ms for Parakeet
  • Accuracy — Higher (Whisper ~5% WER, Parakeet ~1.9% WER)
  • Size — 140 MB (Whisper), 640 MB (Parakeet)
The offline model’s transcript is used for LLM processing.

​
Performance Metrics

After each interaction, listen mode prints: 
STT: 43.2ms
LLM: 87 tok  189 tok/s  TTFT 18ms
TTS: 142ms  0.8x RT
  • STT — Transcription time
  • LLM — Tokens, throughput, time-to-first-token
  • TTS — Synthesis time, real-time factor

​
Stopping Listen Mode

Press Ctrl+C to gracefully stop: 
RCLI — See you next time!

​
Example Session

$ rcli listen

  RCLI — Continuous Voice Mode
  Speak naturally. RCLI listens, acts, and responds.
  Press Ctrl+C to stop.

  Ready.

  /^ ^\
 ( o.o )
  > ^ <
   |_|

Listening for speech...

# User speaks: "open Safari"

  You: open Safari

  /^ ^\
 ( -.- )
  > ^ <
   |_|

Thinking...

  RCLI: Done! Safari is now open.

STT: 39ms
LLM: 12 tok  203 tok/s  TTFT 15ms
TTS: 98ms  0.6x RT

  /^ ^\
 ( o.o )
  > ^ <
   |_|

Listening for speech...

# User speaks: "what's the weather?"

  You: what's the weather?

  RCLI: I don't have access to real-time weather data, 
  but you can ask Siri or check a weather app.

STT: 41ms
LLM: 31 tok  178 tok/s  TTFT 19ms
TTS: 187ms  0.9x RT

^C
  RCLI — See you next time!

​
Troubleshooting

​
No Speech Detected

If RCLI doesn’t respond:
  1. Check microphone permission — System Settings > Privacy & Security > Microphone
  2. Test mic levels — rcli mic-test
  3. Speak louder or closer — VAD threshold is 0.003 RMS

​
Slow Response

If TTFT > 100ms:
  1. Use smaller LLM — rcli models → select Qwen3 0.6B or LFM2 350M
  2. Increase GPU layers — --gpu-layers 99 (default)
  3. Check system load — Close other GPU-heavy apps

​
Incorrect Transcription

If STT accuracy is low:
  1. Upgrade STT model — rcli upgrade-stt (Parakeet TDT, ~1.9% WER)
  2. Speak clearly — Avoid background noise
  3. Check mic quality — Built-in MacBook mic is sufficient

​
Advanced Usage

​
Custom System Prompt

Modify ~/Library/RCLI/config/system_prompt.txt to change LLM behavior: 
You are a helpful voice assistant. Keep responses brief and natural.

​
Action Filtering

Disable noisy actions: 
rcli actions
# Use TUI to toggle actions on/off
# Or edit ~/.rcli/actions.json

​
Multi-Turn Conversations

Listen mode maintains conversation history with sliding window token-budget trimming: 
You: "what's 2 plus 2?"
RCLI: "Four."
You: "multiply that by 10"
RCLI: "Forty."

Build docs developers (and LLMs) love

Get started for freeTalk to us