🎙️ Prompt Engineering for Audio

Getting high-quality results from Gemini and Lyria requires a slightly different approach than text generation. Audio models "listen" for descriptive cues about texture, space, and emotion.

🎲 Soundscape Generation (generate_soundscape)

Best for: Environmental ambience, background textures.

The Gemini 2.0 Live API excels at complex, layered sounds. Use "sensory" adjectives.

• ❌ Bad: "rain in a city"
• ✅ Good: "Heavy rhythmic rain hitting a glass window in a futuristic cyberpunk city, distant low-frequency hovercar hums, neon signs buzzing faintly, binaural spatial audio."

Tip: The "Seamless Loop" trick When generating loops, explicitly tell the model: "Constant background texture with no sudden peaks, suitable for a seamless loop."

---

🗣️ Voice Generation (generate_voice)

Best for: Narration, character dialogue, expressive reading.

The Native Audio model can follow "voice direction" within the prompt.

• Direction: "Read the following in a raspy, old wizard's voice, pausing for dramatic effect after every sentence."
• Emotion: "A nervous, high-pitched voice that cracks slightly when speaking about the monster."

---

🎵 Music & SFX (generate_music / generate_sfx)

Powered by Lyria 3 (Pro/Clip).

Lyria responds well to genre, instrumentation, and BPM.

• Genre-based: "Lo-fi hip hop with a dusty vinyl crackle, mellow electric piano, 85 BPM, relaxing study vibes."
• Instrumentation: "Solo cello playing a melancholic minor-key melody in a large cathedral with long reverb."
• SFX: "A crisp, metallic 'ching' sound effect for a level-up notification, high-pitched and rewarding."