generate_audio

Generate Audio (Speak)

Mentat's voice — text to speech via OpenAI TTS API.

When to Use

• User says "dime", "habla", "speak", "read this aloud", or any request to hear rather than read
• Generating audio briefings, summaries, or narrations
• User explicitly asks for voice/audio output

Mentat Voice Identity

Attribute	Value
Voice	onyx — deep, calm, authoritative
Model	gpt-4o-mini-tts (best Spanish pronunciation + style control)
Speed	1.15 (slightly faster than default)
ES style	ES-LATAM neutro, ritmo firme, sin afectación
EN style	British English, measured cadence

Default instructions (baked into Mentat's character):

Speak in clear Latin American Spanish. Tone: serious but warm, like a trusted strategic advisor. Cadence: deliberate, slightly faster than normal. Emotion: confident and grounded. Accent: neutral Latin American, no Spain. Style: like a senior consultant giving a private briefing.

For English, swap to:

Speak in British English. Tone: serious but warm, like a trusted strategic advisor. Cadence: deliberate, slightly faster than normal. Emotion: confident and grounded. Style: like a senior consultant giving a private briefing.

The voice is consistent across languages. Mentat sounds the same whether speaking Spanish or English — only the language changes, not the character.

Usage

# Basic — speak in Spanish (default)
uv run scripts/generate_audio.py "Análisis completado. Tres puntos clave."

# English
uv run scripts/generate_audio.py "Analysis complete. Three key points." --lang en

# From file
uv run scripts/generate_audio.py --file /tmp/briefing.txt -o /tmp/briefing.mp3

# From stdin (pipe content)
echo "Quick status update" | uv run scripts/generate_audio.py --stdin --lang en

# Custom output path
uv run scripts/generate_audio.py "Hola" -o /tmp/greeting.mp3

# Expressive model with style instructions
uv run scripts/generate_audio.py "Breaking news." -m gpt-4o-mini-tts --instructions "Speak slowly and deliberately, like a trusted advisor"

# Generate comparison samples (all voices)
uv run scripts/generate_audio.py --samples
uv run scripts/generate_audio.py --samples "Custom sample text here"

# Don't auto-play
uv run scripts/generate_audio.py "Silent generation" --no-play

Handler

scripts/generate_audio.py — standalone Python script, only needs requests.

Models

Model	Flag	Best For	Cost
tts-1-hd	-m tts-1-hd (default)	Quality output, briefings	$30/1M chars
tts-1	-m tts-1	Fast drafts, iteration	$15/1M chars
gpt-4o-mini-tts	-m gpt-4o-mini-tts	Expressive, style control	$12/1M chars

Voices (Reference)

Voice	Character	Mentat fit
onyx	Deep, authoritative	Default — matches Mentat's analytical directness
echo	Clear, direct	Backup — slightly lighter
fable	Warm, British-leaning	Good for EN-UK content
ash	Warm, conversational	Too casual
coral	Warm, friendly	Too soft
nova	Energetic	Too bright
sage	Wise, calm	Alternative if onyx is too heavy
shimmer	Bright, optimistic	Doesn't fit
alloy	Neutral, balanced	Generic
ballad	Storytelling	Too performative

Rules

• Always run from hive/ directory
• Default output: /tmp/mentat_speech.mp3
• Auto-plays on macOS after generating (use --no-play to suppress)
• Max 4096 chars per request (API limit) — longer texts get truncated with warning
• For long content, split into logical chunks and generate separately
• Voice selection is an identity decision — don't change without user request

Workflow: Spoken Response

1. Prepare text (clean markdown, remove code blocks, keep prose)
2. Detect language (default ES, switch to EN if content is English)
3. Generate with Mentat's voice
4. Auto-play
5. Save path for reference

Anti-patterns

Don't	Why	Do Instead
Read raw markdown aloud	Sounds robotic with headers/bullets	Convert to natural prose first
Switch voices mid-session	Breaks identity consistency	Always use Mentat's voice unless sampling
Generate > 4096 chars	API truncates silently	Split into chunks
Use tts-1 for final output	Audibly lower quality	Use tts-1-hd (default)