Say + FFmpeg TTS Pipeline

Use say (macOS native TTS) + ffmpeg to generate Opus voice messages for Telegram/Matrix.

Why not just say?

• Telegram/Matrix require Opus codec voice messages
• say outputs AIFF/m4a; must convert to .ogg (Opus) before sending
• Telegram accepts: OGG/MP3/M4A as voice — but Opus OGG is the native format

Workflow

say -v "\x3Cvoice>" -o \x3Ctmpdir>/\x3Cname>.aiff "\x3Ctext>"
ffmpeg -i \x3Ctmpdir>/\x3Cname>.aiff -acodec libopus \x3Ctmpdir>/\x3Cname>.ogg -y

Send with message tool:

{
  "action": "send",
  "channel": "telegram",
  "media": "\x3Ctmpdir>/\x3Cname>.ogg",
  "asVoice": true,
  "target": "\x3Cchat_id>"
}

Recommended workspace directory

~/.openclaw/workspace/tmp/audio/

(Whitelist this path in exec permissions for faster approval)

Voice selection

Use say -v '?' to list available voices. Notable ones:

• Trinoids — robotic/electronic voice (popular for bots)
• Samantha — warm US female voice
• Alex — US male voice
• Fred — neutral US male voice
• Karen — Australian female voice

Note: pass just the voice name (e.g. "Trinoids"), not the full en_US suffix.

Example: send a hello voice message

VOICE="Trinoids"
TEXT="Hello!"
DIR="$HOME/.openclaw/workspace/tmp/audio"
mkdir -p "$DIR"

say -v "$VOICE" -o "$DIR/hello.aiff" "$TEXT"
ffmpeg -i "$DIR/hello.aiff" -acodec libopus "$DIR/hello.ogg" -y

# Then send via message tool with asVoice: true

Format notes

• Input to ffmpeg: AIFF (.aiff) works reliably; avoid .m4a with say
• Output: Opus in Ogg container (libopus codec) — required for Telegram voice messages
• Telegram sendVoice accepts: OGG, MP3, M4A — but native is Opus OGG
• Sample rate: say outputs 24kHz AIFF; ffmpeg re-encodes to Opus at 24kHz

Integration with OpenClaw TTS

OpenClaw's built-in messages.tts only supports: ElevenLabs, Microsoft Edge, MiniMax, OpenAI.

This say+ffmpeg pipeline is a workaround for local-only TTS without API keys or cloud services. It's not auto-triggered by OpenClaw — call it manually via exec + message tool.

Language Detection → Voice Mapping

When responding to a voice message, detect the language from the STT output (Parakeet auto-detects). Then pick the matching say voice using i18n locale codes.

Finding voices by language:

say -v '?' 2>&1 | grep -E "cs_CZ|en_US|de_DE|fr_FR|it_IT|es_ES"

Language → voice selection priority:

1. Use \x3Cvoice> (Premium) if available
2. Fall back to \x3Cvoice> (Enhanced) if available
3. Fall back to base \x3Cvoice> name
4. Never use a voice that doesn't match the language

Key: Always use just the voice name (e.g. "Trinoids", "Zuzana"), not the full locale suffix. The locale suffix in say -v '?' output is for grepping/identification only.

Example workflow:

LANG="cs_CZ"
# Find best available voice for this language (Premium > Enhanced > base)
VOICE=$(say -v '?' 2>&1 | grep "$LANG" | head -3 | awk '{print $1}' | sed -n '1p')
say -v "$VOICE" -o reply.aiff "Česká odpověď"
ffmpeg -i reply.aiff -acodec libopus reply.ogg -y

TODOs

• Detect language from STT transcription and auto-select appropriate say voice
• Explore integrating into OpenClaw via custom TTS provider plugin
• Investigate if OpenClaw supports post-processing TTS output via a hook
• Test Matrix channel voice message format compatibility