cartesia tts script

cartesia-tts.py

import os
import requests
from dotenv import load_dotenv
import numpy as np
import soundfile as sf

# Load environment variables from .env file
load_dotenv()

# Base URL for Cartesia API
CARTESIA_TTS_URL = "https://api.cartesia.ai/tts/bytes"


def get_headers():
    """Return headers for Cartesia API."""
    return {
        "Cartesia-Version": os.getenv("CARTESIA_VERSION", "2024-01-01"),
        "Authorization": f"Bearer {os.getenv('CARTESIA_API_TOKEN')}",
        "Content-Type": "application/json"
    }


def build_payload(model_id: str, transcript: str, voice_id: str, mode: str = "id"):
    """Build the JSON payload for TTS request."""
    return {
        "model_id": model_id,
        "transcript": transcript,
        "voice": {
            "mode": mode,
            "id": voice_id
        },
        "language": "en",
        "generation_config": {
            "volume": 1,
            "speed": 1,
            "emotion": "neutral"
        },
        "output_format": {
            "container": "raw",
            "encoding": "pcm_f32le",
            "sample_rate": 8000
        },
        "save": False,
        "speed": "normal"
    }


def synthesize_speech(model_id: str, transcript: str, voice_id: str, mode: str = "id"):
    """Call Cartesia API and return the response."""
    headers = get_headers()
    payload = build_payload(model_id, transcript, voice_id, mode)

    response = requests.post(CARTESIA_TTS_URL, json=payload, headers=headers)
    response.raise_for_status()  # Raise error for bad status codes

    return response.content


def pcm_f32le_raw_to_wav(raw: bytes, path: str, sample_rate: int = 8000, channels: int = 1) -> None:
    # Interpret raw bytes as float32 little-endian and shape (N, channels)
    audio = np.frombuffer(raw, dtype="<f4")
    if channels > 1:
        audio = audio.reshape(-1, channels)
    sf.write(path, audio, samplerate=sample_rate, subtype="FLOAT")  # Writes IEEE float WAV


if __name__ == "__main__":
    # Example usage
    model = "sonic-3"
    voice = "cc00e582-ed66-4004-8336-0175b85c85f6"
    text = "You need to wrap the raw samples into a WAV. Two ways:"
    output_path = 'speech4.wav'

    result = synthesize_speech(model, text, voice)
    pcm_f32le_raw_to_wav(result, output_path)