tts-generation
Konversi naskah acara radio menjadi audio ucapan dengan efek telepon pada suara koresponden menggunakan Interactions API.
npx skills add https://github.com/google-gemini/gemini-managed-agents-templates --skill tts-generationTTS Generation
Convert the radio show script into speech audio using the Gemini TTS model via the Interactions API. Apply a telephone bandpass filter to correspondent voices so they sound like phone call-ins, while keeping the host's voice clean studio-quality.
Embedded Script
python3 skills/tts-generation/scripts/generate_tts.py --workspace ./workspace
Arguments
| Argument | Default | Description |
|---|---|---|
--workspace | workspace | Root workspace directory |
--workers | 8 | Max parallel TTS worker threads |
What it does
- Reads the script from
{workspace}/data/script.md. - Parses it into individual
(speaker, text)turns and assigns voices. - Generates TTS for all turns in parallel using a thread pool (default 8 workers).
- Retries each failed turn up to 3 times with exponential backoff.
- Applies an ffmpeg telephone bandpass filter (300Hz–3.4kHz) to correspondent voices.
- Keeps the host (Paul) audio clean and unfiltered.
- Concatenates all segments in original script order into a single WAV.
Dependencies
google-genai(>= 2.0.0)ffmpeg(system)
API Details
Uses the Interactions API with single-speaker TTS — no multi-speaker workaround needed.
Voice Assignment
Voices are assigned dynamically based on [Male] / [Female] gender tags in the script:
| Speaker | Voice | Audio Treatment |
|---|---|---|
| Paul (host) | Puck | Clean — no filter |
Caller [Female] — 1st | Kore | Telephone filter |
Caller [Female] — 2nd | Aoede | Telephone filter |
Caller [Male] — 1st | Charon | Telephone filter |
Caller [Male] — 2nd | Fenrir | Telephone filter |
Voices cycle round-robin if there are more callers than available voices. Accent tags ([Accent: Irish], etc.) are injected into the TTS prompt to influence pronunciation.
Telephone Filter
Applied via ffmpeg to correspondent audio segments:
highpass=f=300, lowpass=f=3400, acompressor, volume=1.5
This simulates the standard telephone bandwidth (300Hz–3.4kHz) and adds compression to mimic phone codec dynamics.
Output
- Primary output:
{workspace}/audio/speech/speech.wav - Format: WAV, 24kHz, 16-bit PCM, mono
- Intermediate segments:
{workspace}/audio/speech/segments/turn_*.wav