Gemini Tts

Gemini Text-to-Speech

Generate natural, expressive speech from text using Google Gemini TTS. Supports 70+ languages, 30 prebuilt voices, and style control via natural language prompts and audio tags.

Requirements: Node.js 18+, ffmpeg, @google/genai npm package, GEMINI_API_KEY env var.

Available Scripts

• scripts/generate-tts.mjs — All-in-one TTS generator with single-request consistency, silence-based splitting, and automatic trimming.

Quick Start

Single text to speech

npm install @google/genai
GEMINI_API_KEY=your-key node scripts/generate-tts.mjs \
  --text "Hello, welcome to our product!" \
  --output-dir public/vo \
  --voice Kore

Multi-segment voiceover (consistent voice)

Create a segments.json:

[
  { "id": "intro", "text": "Welcome to our product launch." },
  { "id": "features", "text": "Here are the three features you will love." },
  { "id": "close", "text": "Get started today at our website." }
]

Run:

GEMINI_API_KEY=your-key node scripts/generate-tts.mjs \
  --segments segments.json \
  --output-dir public/vo \
  --voice Kore \
  --director-notes "Tone: Warm, confident, inspiring. Style: Professional narrator with a vocal smile. Accent: Neutral American English. Pacing: Moderate."

Output (JSON to stdout):

{
  "success": true,
  "segments": [
    { "id": "intro", "file": "public/vo/vo-intro.mp3", "duration": 3.25 },
    { "id": "features", "file": "public/vo/vo-features.mp3", "duration": 4.10 },
    { "id": "close", "file": "public/vo/vo-close.mp3", "duration": 2.80 }
  ],
  "totalDuration": 10.15,
  "fullAudioFile": "public/vo/vo-full.wav"
}

Script Options

Exit Codes

How It Works (Multi-Segment)

Gemini TTS does NOT guarantee voice consistency across separate API requests. The script solves this by:

1. Combining all segments into ONE prompt with [3 second pause] markers between them
2. Generating a single continuous audio file in one API call
3. Detecting silence gaps with ffmpeg silencedetect using adaptive thresholds (starts at noise=-35dB, d=1.5s, progressively relaxes). If an exact match is found, uses it directly. If a threshold jump occurs (too few → too many), regenerates audio with faster speech pacing to make inter-segment pauses more distinct, then retries detection.
4. Splitting at silence midpoints into individual segment files
5. Trimming residual silence padding (~1.5s per side) from each segment, keeping 0.15s breathing room

Models

Voices

30 prebuilt voices:

Director's Notes (--director-notes)

Performance guidance for the selected voice. Define only what matters — overspecifying limits the model's creativity and can worsen results.

The three most common directions are Style, Pacing, and Accent, but you can include any custom instructions relevant to the performance.

• Style — Emotional quality and delivery character. Be descriptive: "Infectious enthusiasm — the listener should feel part of something exciting" works better than just "energetic". You can layer characteristics and use voiceover industry terms like "vocal smile".
• Pacing — Overall speed and rhythm variation. Can range from "Speak as fast as possible" to "Incredibly slow and liquid, words bleed into each other, zero urgency".
• Accent — Be specific. "British English as heard in Croydon, England" works much better than "British accent".

--director-notes "Style: Enthusiastic and sassy GenZ beauty YouTuber. Pacing: Extremely fast, rapid delivery like short-form video influencers. Accent: Southern California valley girl from Laguna Beach."

Audio Tags

Inline [brackets] in segment text for per-word control. Override director-notes locally.

[whispers] [shouting] [excited] [serious] [short pause] [laughs] [very fast] [very slow] [sarcastic] [curious]

Creative tags work too: [like dracula], [whispering conspiratorially]

Example: "[excited] This changes everything! [short pause] Let me show you how."

References

• Prompting Guide — Advanced prompt engineering for style, emotion, and accent