lettabot/docs/voice.md

Voice

LettaBot has full voice support: it can receive voice messages (transcribed to text) and reply with voice memos (generated via TTS). Both features work across Telegram, WhatsApp, Signal, Discord, and Slack.

Voice Transcription (Receiving Voice Messages)

When a user sends a voice message, LettaBot downloads the audio, transcribes it via the configured provider, and delivers the text to the agent with a [Voice message]: prefix.

Providers

OpenAI Whisper (default):

transcription:
  provider: openai
  apiKey: sk-...       # Optional: falls back to OPENAI_API_KEY env var
  model: whisper-1     # Default

Mistral Voxtral (faster, lower cost):

transcription:
  provider: mistral
  apiKey: ...          # Optional: falls back to MISTRAL_API_KEY env var
  model: voxtral-mini-latest  # Default

Or configure via environment variables alone:

# OpenAI (default provider when no config is set)
export OPENAI_API_KEY=sk-...

# Mistral (requires provider to be set in config)
export MISTRAL_API_KEY=...

If no API key is configured, users who send voice messages will receive an error message with a setup link.

Supported Audio Formats

These formats are sent directly to the transcription API (some with a filename remap):

flac, m4a, mp3, mp4, mpeg, mpga, oga, ogg, opus, wav, webm

These formats are automatically converted to MP3 via ffmpeg (if installed):

aac, amr, caf, 3gp, 3gpp

Files over 20MB are automatically split into 10-minute chunks before transcription.

Channel Support

Channel	Format received	Notes
Telegram	OGG/Opus	Native voice messages
WhatsApp	OGG/Opus	Push-to-talk voice messages
Signal	Various	Voice attachments
Discord	Various	Audio file attachments
Slack	Various	Audio file uploads

Voice Memos (Sending Voice Notes)

The agent can reply with voice notes using the <voice> directive. The text is sent to a TTS provider, converted to OGG Opus audio, and delivered as a native voice bubble (on Telegram and WhatsApp) or a playable audio attachment (on Discord and Slack).

How It Works

The agent includes a <voice> tag in its response:

<actions>
  <voice>Hey, here's a quick update on that thing we discussed.</voice>
</actions>

This can be combined with text -- anything after the </actions> block is sent as a normal message alongside the voice note:

<actions>
  <voice>Here's the summary as audio.</voice>
</actions>
And here it is in text form too!

See directives.md for the full directive reference.

Providers

ElevenLabs (default):

tts:
  provider: elevenlabs
  apiKey: sk_475a...                    # Or ELEVENLABS_API_KEY env var
  voiceId: onwK4e9ZLuTAKqWW03F9         # Or ELEVENLABS_VOICE_ID env var
  model: eleven_multilingual_v2         # Or ELEVENLABS_MODEL_ID

Browse voices at elevenlabs.io/voice-library.

OpenAI:

tts:
  provider: openai
  apiKey: sk-...                        # Or OPENAI_API_KEY env var
  voiceId: alloy                        # Or OPENAI_TTS_VOICE (options: alloy, echo, fable, onyx, nova, shimmer)
  model: tts-1                          # Or OPENAI_TTS_MODEL (use tts-1-hd for higher quality)

Channel Support

Channel	Delivery	Notes
Telegram	Native voice bubble	Falls back to audio file if user has voice message privacy enabled (Telegram Premium). Users can allow via Settings > Privacy and Security > Voice Messages.
WhatsApp	Native voice bubble	Sent with push-to-talk (ptt: true) for native rendering.
Discord	Audio attachment	Playable inline.
Slack	Audio attachment	Playable inline.
Signal	Audio attachment	Sent as a file attachment.

When to Use Voice

• User sent a voice message and a voice reply feels natural
• User explicitly asks for a voice/audio response
• Short, conversational responses (under ~30 seconds of speech)

When NOT to Use Voice

• Code snippets, file paths, URLs, or structured data -- these should be text
• Long responses (keep voice under ~30 seconds)
• When the user has indicated a preference for text

CLI Tools

lettabot-tts

Generate audio from the command line:

lettabot-tts "Hello, this is a test"           # Outputs file path to stdout
lettabot-tts "Hello" /tmp/output.ogg            # Explicit output path

Output files are written to data/outbound/ by default and auto-cleaned after 1 hour.

lettabot-message --voice

Send a voice note from a background task (heartbeat, cron):

# Generate + send in one step
OUTPUT=$(lettabot-tts "Your message here") || exit 1
lettabot-message send --file "$OUTPUT" --voice

# Send to a specific channel
lettabot-message send --file "$OUTPUT" --voice --channel telegram --chat 123456

Environment Variable Reference

Variable	Description	Default
Transcription
OPENAI_API_KEY	OpenAI API key (Whisper transcription + OpenAI TTS)	--
MISTRAL_API_KEY	Mistral API key (Voxtral transcription)	--
TRANSCRIPTION_MODEL	Override transcription model	whisper-1 / voxtral-mini-latest
Text-to-Speech
TTS_PROVIDER	TTS backend	elevenlabs
ELEVENLABS_API_KEY	ElevenLabs API key	--
ELEVENLABS_VOICE_ID	ElevenLabs voice ID	onwK4e9ZLuTAKqWW03F9
ELEVENLABS_MODEL_ID	ElevenLabs model	eleven_multilingual_v2
OPENAI_TTS_VOICE	OpenAI TTS voice name	alloy
OPENAI_TTS_MODEL	OpenAI TTS model	tts-1

All environment variables can be overridden by the equivalent YAML config fields (see above).

Troubleshooting

Voice messages not transcribing

1. Check that an API key is configured -- either in lettabot.yaml under transcription.apiKey or via the OPENAI_API_KEY / MISTRAL_API_KEY environment variable
2. Check the logs for transcription errors
3. If using an unsupported audio format, install ffmpeg for automatic conversion

Voice memos not generating

1. Check that a TTS provider is configured -- either in lettabot.yaml under tts or via ELEVENLABS_API_KEY / OPENAI_API_KEY
2. Check that jq and curl are installed (required by the lettabot-tts script)
3. Check logs for TTS API errors (HTTP status codes, rate limits)

Telegram voice privacy

If the bot sends audio files instead of voice bubbles on Telegram, the recipient has voice message privacy enabled (Telegram Premium feature). They can allow voice messages via Settings > Privacy and Security > Voice Messages.