Text-to-Speech (TTS)

TTS Landscape (2024/2025)

Modern TTS has moved far beyond robotic, concatenated synthesis. Neural TTS models now generate speech that is often indistinguishable from human speakers, capturing emotion, prosody, and breath.

OpenAI TTS API

Best for general-purpose applications needing high quality at a reasonable usage cost.

from pathlib import Path
from openai import OpenAI

client = OpenAI()

speech_file_path = Path(__file__).parent / "speech.mp3"

response = client.audio.speech.create(
    model="tts-1",           # or "tts-1-hd" for higher quality
    voice="alloy",           # Options: alloy, echo, fable, onyx, nova, shimmer
    input="The quick brown fox jumps over the lazy dog."
)

response.stream_to_file(speech_file_path)
print(f"Saved to {speech_file_path}")

ElevenLabs (Voice Cloning)

The standard for voice cloning and emotional range. Requires an API key.

import os
from elevenlabs.client import ElevenLabs
from elevenlabs import save

client = ElevenLabs(
    api_key="YOUR_API_KEY",
)

audio = client.generate(
    text="Hello! This is a realistic AI voice generated by ElevenLabs.",
    voice="Rachel",
    model="eleven_multilingual_v2"
)

save(audio, "elevenlabs_output.mp3")

Micro-Service: MeloTTS (Open Source)

Run high-quality TTS locally or on your own server. MeloTTS is fast enough for CPU inference.

1. Installation

pip install melotts
# Requires mecab for some languages
python -m unidic download

2. Python Usage

from melo.api import TTS

# Speed is adjustable
speed = 1.0
device = 'cpu' # or 'cuda:0'

model = TTS(language='EN', device=device)
speaker_ids = model.hps.data.spk2id

# Use American English accent
output_path = 'melo_en.wav'
model.tts_to_file("Hello, this is running locally on your machine.", speaker_ids['EN-US'], output_path, speed=speed)

Real-time Streaming Pattern

For voice assistants, waiting for the full audio file is too slow. Use streaming to play audio chunks as they arrive.

# Example using OpenAI Streaming
response = client.audio.speech.create(
    model="tts-1",
    voice="alloy",
    input="This text is being streamed in real-time.",
)

# Stream to a file or audio output
response.stream_to_file("output.mp3")

# Ideally, pipe 'response.content' chunks to an audio player like PyAudio
# for true low-latency playback.

Comparison Matrix

Related Topics