cog-voice
Skill4 starsChanged 44 days agoMITView on GitHub
Dual-modal communication via the Mod³ TTS server — when to speak vs write, non-blocking speech discipline, reading voice metrics, anti-patterns. Use when voice output is wanted alongside text.
- Sends data out
In this repositoryOne other file it gives its agents.
- bodySkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 35
Write the structured result as text — code, data, analysis
- Distinctive: line 40
Use speech for code, file paths, or anything the user needs to copy
- Distinctive: line 122
Ellipses: add pauses: "well... let me think..."
- Distinctive: line 128
Narrating your actions in speech: Don't say "I'm going to read the file now." Just do it.
- Distinctive: line 129
Long monologues: Keep speech concise. If it's more than 3-4 sentences, it should probably be text.
- Distinctive: line 132
Speaking code: Never. Code is text. Always.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Voice — Dual-Modal Communication via Mod³
You have a voice. Use it intentionally.
The Two Channels
You now have two output channels with different characteristics:
| Channel | Persistence | Latency | Best for |
| Text | Permanent (visible in conversation) | Instant | Code, structured data, decisions, diffs, anything the user will reference later |
| Voice | Ephemeral (heard once, not in transcript) | ~0.5s TTFA | Context, thinking out loud, status updates, conversational responses, emotional tone |
The rule: Voice carries the ephemeral. Text carries the persistent. Don't duplicate — use each channel for what it's good at.
Non-Blocking Speech
speak() returns immediately. Audio plays in the background while you continue working. This is the core capability — you can talk and act simultaneously.
speak("Looking into that now, give me a second.") → returns job_id instantly
# ... do the actual work here while user hears you ...
# ... write the structured result as text ...
speech_status(job_id) → check metrics if neededDo this:
- Speak a brief orientation ("Let me check that" / "Found it" / "Here's what I see") while performing the action
- Write the structured result as text — code, data, analysis
- The user hears your intent and sees your output simultaneously
Don't do this:
- Speak the same content you're about to write (redundant)
- Use speech for code, file paths, or anything the user needs to copy
- Block on speech_status() unless you need the metrics for diagnostic purposes
When to Speak
Speak when:
- Acknowledging a request before starting work
- Giving a brief summary while detailed output renders as text
- Responding to conversational/emotional content where tone matters
- Explaining your reasoning while the code writes itself
- Something surprising happens (errors, discoveries, results)
Write when:
- Outputting code, configs, structured data
- Anything the user will reference, copy, or search for later
- Detailed technical analysis
- File paths, URLs, commands
Both simultaneously (the power move):
- Speak the high-level context while writing the detailed output
- "I found three issues" (voice) + the actual issue list (text)
- "Deploying now" (voice) + the deployment log (text)
Voice Selection
Cog's voice is eng_uk_m_davids on chatterbox-turbo. Recorded at
cog:mem/semantic/insights/cog-voice-preference.cog.md. Use it unless the
moment calls for something else. The old bm_lewis default was inherited
plumbing, not a choice.
Multiple engines are available. Pick based on the moment:
- Default (eng_uk_m_davids): chatterbox-turbo, British male. Warm RTF ~1.3-1.8, faster than realtime.
- Kokoro (bm_lewis, af_heart, ...): fastest TTFA (~58ms), supports speed + ALL CAPS emphasis. Use when latency dominates.
- Voxtral voices (casual_male, neutral_female, etc.): best raw quality, zero control surfaces, brutal ~60s cold load. Use sparingly.
- Chatterbox (full, not turbo): the ONLY engine that honors
emotion. Turbo silently ignores it. One stock voice unless you passref_audio.
chatterbox-turbo also carries 59 clone profiles — attenborough, sagan,
einstein, frasier, her, kronk, plus regional English sets. Call
list_voices to see them (fixed 2026-08-19; older builds hid all but one).
Rendering to a File vs Playing Aloud
Two surfaces. Confusing them is the most common mistake:
| Path | Effect | Use when |
MCP speak | Plays out of darkstar's physical speakers | Chaz is at the machine |
HTTP POST /v1/synthesize | Returns bytes, silent | Sending audio over Discord/Telegram |
From a chat platform, speak fires audio into an empty room. Use the HTTP path:
curl -s -X POST http://127.0.0.1:7860/v1/synthesize \
-H 'Content-Type: application/json' \
-d '{"text":"...","voice":"eng_uk_m_davids","format":"ogg"}' \
-D /tmp/hdr.txt -o /tmp/out.ogg
ffmpeg -y -loglevel error -i /tmp/out.ogg -c:a libmp3lame -b:a 128k /tmp/out.mp3Transcode to mp3 — Discord will not inline .ogg. Metrics return in
x-mod3-* headers.
Reading the Metrics
speech_status() returns structured metrics. Key signals:
- underruns > 0: Audio had gaps. GPU was under load. Consider shorter utterances or simpler model.
- ttfa_sec > 2.0: First audio was slow. Model may be cold-loading. Subsequent calls will be faster.
- overall_rtf < 1.0: Generation slower than playback. Expect gaps. Switch to a lighter model (Kokoro).
- buffer.min_samples = 0: Buffer emptied during playback. Audio likely stuttered.
You don't need to check metrics on every call. Check when:
- The user reports audio issues
- You want to diagnose why something sounded wrong
- You're experimenting with different models/speeds
Text Formatting for Prosody
With Kokoro voices, text formatting affects speech:
- ALL CAPS adds emphasis: "how are YOU doing" stresses "you"
- Ellipses add pauses: "well... let me think..."
- Punctuation shapes intonation: questions rise, exclamations emphasize
- Short sentences produce cleaner prosody than long compound ones
Anti-Patterns
- Narrating your actions in speech: Don't say "I'm going to read the file now." Just do it. Speak only when the user benefits from hearing it.
- Long monologues: Keep speech concise. If it's more than 3-4 sentences, it should probably be text.
- Waiting for speech to finish: The whole point is non-blocking. Don't call speech_status() in a loop.
- Using speech as a crutch for filler: "Let me think about that..." is fine once. Don't stall with voice.
- Speaking code: Never. Code is text. Always.
The Philosophy
Voice and text are not redundant channels — they're complementary modalities with different information densities and persistence profiles. Using both simultaneously is not showing off; it's the natural way to communicate when you have both available. A human explaining code talks while pointing at the screen. You speak while writing. Same instinct, same efficiency.
The user hears your intent. The user sees your output. Both arrive together. That's dual-modal communication.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

