awesome-generative-ai

Text-to-Speech (TTS) Models

Curated collection of high-quality open-source TTS models and toolkits for research, production, and multi-language synthesis.

---

Table of Contents

• High-Fidelity Models
• Fast and Efficient Models
• Vocoders
• Notable TTS Projects
• Additional TTS Models
• New Additions (Curated)
• Selection Guide
• Additional Resources

---

High-Fidelity Models

Tortoise-TTS

• Type: High-fidelity voice cloning and generation
• Features: Prompt tuning, long-form speech
• Requirements: GPU, ~6GB memory recommended
• Best for: High-quality voice synthesis

VoiceCraft

• Type: Zero-shot speech editing and text-to-speech
• Features: Speech editing in the wild, speaker conditioning
• Best for: High-quality speech editing and synthesis

VITS (Variational Inference TTS)

• Type: End-to-end model combining Tacotron2 + HiFi-GAN
• Features: High-quality synthesis, fewer artifacts
• Performance: Fast training and inference
• Best for: Research and production

Coqui TTS

• Type: Modular, multilingual toolkit
• Features: Easy training, fine-tuning, pretrained models
• Languages: English, German, French, and more
• Best for: Production-ready applications

Bark

• Type: Multilingual text-to-audio model
• Features: High-quality speech synthesis
• Languages: Multiple languages
• Best for: Creative audio generation

Maha TTS

• Type: Multilingual TTS system
• Features: Indian languages support
• Languages: Hindi, English, and regional languages
• Best for: Indian language applications

MMS (Massively Multilingual Speech)

• Type: Multilingual speech recognition and synthesis
• Features: 1000+ languages support
• Languages: Extensive language coverage
• Best for: Global applications

Vall-E X

• Type: Multilingual TTS with voice cloning
• Features: Cross-lingual voice cloning
• Languages: Multiple languages
• Best for: Multilingual voice synthesis

StyleTTS2

• Type: Style-aware TTS system
• Features: Style control and transfer
• Performance: High-quality synthesis
• Best for: Expressive speech generation

SeamlessM4T

• Type: Multimodal multilingual model
• Features: Speech-to-speech translation
• Languages: 100+ languages
• Best for: Real-time translation

---

Fast and Efficient Models

Tacotron 2

• Type: Two-stage pipeline (spectrogram + vocoder)
• Features: Realistic prosody and intonation
• Framework: TensorFlow
• Best for: Research and learning

FastSpeech 2

• Type: Non-autoregressive TTS
• Features: High speed and stability
• Vocoders: HiFi-GAN, WaveGlow compatible
• Best for: Real-time applications

Glow-TTS

• Type: Flow-based TTS architecture
• Features: High-performance, parallelizable
• Best for: Fast inference scenarios

KittenTTS

• Type: Ultra-lightweight neural TTS
• Features: 15M parameters, <25MB model size
• Performance: CPU-optimized, no GPU required
• Voices: 8 premium voice options (male/female variants)
• Best for: Edge deployment, mobile applications

Piper

• Type: Fast, local neural text-to-speech system
• Features: Lightweight deployment, high-quality synthesis
• Performance: Optimized for real-time applications
• Languages: Multiple language support
• Note: Development moved to OHF-Voice/piper1-gpl
• Best for: Local deployment, privacy-focused applications

---

Vocoders

Vocoders convert spectrograms to audio waveforms.

HiFi-GAN

• Type: Fast, high-quality vocoder
• Features: Efficient waveform generation
• Best for: Production TTS systems

WaveGlow

• Type: Real-time waveform generator
• Features: NVIDIA-optimized
• Best for: GPU-accelerated synthesis

MelGAN

• Type: GAN-based vocoder
• Features: Adversarial training
• Best for: High-quality audio generation

---

Notable TTS Projects

Fish-Speech

• Type: Real-time streaming TTS system
• Features: High-fidelity, low-latency synthesis
• Demo: https://speech.fish.audio/samples/
• Website: https://speech.fish.audio/

Kokoro

• Type: Multilingual expressive TTS
• Features: Multi-speaker, style control
• Languages: Japanese and English
• PyPI: https://pypi.org/project/kokoro/
• Model: https://huggingface.co/hexgrad/Kokoro
• Demo: https://huggingface.co/spaces/hexgrad/kokoro-tts

Llasa-TTS

• Type: Ultra-fast multilingual TTS engine
• Features: Streaming, natural prosody
• Demo: https://huggingface.co/spaces/srinivasanbalasubramani/llasa-tts

Spark-TTS

• Type: Modular, real-time neural TTS
• Features: Latency-optimized
• Demo: https://sparkaudio.github.io/spark-tts/

Qwen3-TTS

• Type: Large-scale multilingual TTS model
• Features: High-quality synthesis, controllable voices
• Best for: Research and production TTS

ComfyUI-Qwen-TTS

• Type: ComfyUI node / workflow integration for Qwen TTS
• Features: Generate TTS audio directly inside ComfyUI graphs
• Best for: TTS pipelines built in ComfyUI

VITS2

• Type: Tacotron-style neural TTS
• Features: Real-time performance, lightweight
• Architecture: Fast and modular
• Synthesis: Phoneme-based via espeak-ng

Index-TTS

• Type: Neural text-to-speech system
• Features: Training and inference pipeline
• Best for: Research and experimentation

Chatterbox

• Type: Neural text-to-speech model
• Features: Open-source TTS model and demos
• Best for: Prototyping expressive speech

ChatTTS

• Type: Conversational TTS model
• Features: Dialogue-focused prosody and expressiveness
• Best for: Chat-style TTS and assistant voices

FireRedTTS2

• Type: Neural text-to-speech system
• Features: Training and inference codebase
• Best for: Model exploration and benchmarks

Genie-TTS

• Type: Text-to-speech toolkit
• Features: Inference tooling and model assets
• Best for: Quick demos and experiments

Supertonic

• Type: Text-to-speech model and toolkit
• Features: Repository with model and utilities
• Best for: Research and evaluation

Granite Speech Models

• Type: Enterprise-focused speech model collection
• Features: Granite family speech model assets and usage examples
• Best for: IBM Granite speech workflows and enterprise prototyping

TADA-TTS

• Type: Neural text-to-speech project
• Features: Open-source codebase for training and inference
• Best for: TTS experimentation and custom speech pipelines

---

Additional TTS Models

Extension Models

• XTTSv2 - Enhanced XTTS implementation
• MARS5 - Multilingual TTS system
• F5-TTS - High-quality diffusion/flow TTS (see curated entry below)
• Parler TTS - Hugging Face TTS
• OpenVoice - Open-source voice cloning
• OpenVoice V2 - Enhanced OpenVoice
• MeloTTS - Multilingual text-to-speech toolkit by MyShell
• Irodori-TTS - Neural text-to-speech project and tooling
• kani-tts-2 - TTS model and inference toolkit
• FlashLabs-Chroma - Speech synthesis project and tooling
• DIA - Neural TTS framework
• Higgs Audio - Open-source TTS model and toolkit
• Soprano - TTS model and training toolkit
• CosyVoice - Conversational TTS
• VoxCPM - Open-source TTS model
• GLM-TTS - Zero-shot controllable TTS with emotion and style tokens
• GPT-SoVITS - GPT-based voice synthesis
• Piper TTS - Lightweight TTS engine
• Kimi Audio 7B Instruct - Large-scale audio model
• ACE-Step - Advanced TTS framework

---

New Additions (Curated)

E2-TTS (e2-tts-pytorch)

• Type: Fully non-autoregressive end-to-end TTS (PyTorch)
• Features: Zero-shot style-capable synthesis, multistream transformer conditioning, simple training and inference API
• Requirements: PyTorch (see repo for CUDA/CPU options)
• Best for: Research experiments and fast end-to-end TTS prototyping

F5-TTS

• Type: Diffusion/flow-based TTS mixing ConvNeXt V2 + Flow Matching
• Features: High-quality, faster training/inference, Gradio/CLI apps, Triton/TensorRT runtime guides
• Requirements: GPU recommended for best performance
• Best for: High-quality synthesis and production benchmarks

Delayed Streams Modeling / Kyutai

• Type: Collection for streaming STT and TTS (Kyutai models)
• Features: Streaming/low-latency TTS examples, PyTorch + Rust + MLX implementations, Colab demos
• Requirements: Varies by implementation (PyTorch, Rust, or Apple MLX)
• Best for: Real-time/streaming TTS systems and production servers

Chatterbox Fine-Tuning Kit

• Type: Fine-tuning and inference kit for Chatterbox TTS (Standard and Turbo modes)
• Features: Automated tokenizer merging, preprocessing pipeline, Turbo mode for faster adaptation, inference scripts with VAD
• Requirements: Python 3.8+, GPU recommended; run setup.py to download base models
• Best for: Fine-tuning voice cloning and multi-language adaptation

Emotional VITS

• Type: VITS variant with emotion conditioning
• Features: Emotion embedding extraction from reference audio, emotion-controllable synthesis
• Requirements: Python 3.6+, preprocessing and monotonic alignment build steps
• Best for: Emotion-aware TTS and expressive voice synthesis research

Matcha-TTS

• Type: Fast non-autoregressive TTS using conditional flow matching
• Features: ICASSP implementation, fast synthesis, ONNX export and runtime support, CLI and Gradio app
• Requirements: PyTorch 2.0+, optional ONNX runtime for export/inference
• Best for: Low-latency, exportable TTS pipelines and production-friendly deployment

---

Selection Guide

Use Case	Recommended Model	Why
High-quality synthesis	Tortoise-TTS	Best audio quality
Production deployment	Coqui TTS	Modular and well-documented
Real-time applications	FastSpeech 2	Fast inference
Research projects	VITS	End-to-end and efficient
Multilingual support	MMS, Vall-E X	Extensive language coverage
Streaming applications	Llasa-TTS	Ultra-fast, streaming
Lightweight deployment	VITS2	Small footprint
Ultra-lightweight/Edge	KittenTTS	<25MB, CPU-only
Local/Privacy-focused	Piper	Fast local synthesis
Voice cloning	OpenVoice, Bark	High-fidelity cloning

---

Voice Apps & Utilities

Voice

• Type: Minimalistic audiobook player (Android)
• Focus: Local playback and simple library management
• Best for: Lightweight audiobook listening apps

---

Additional Resources

• Voice Cloning - Voice synthesis and cloning techniques
• STT Models - Speech-to-text recognition
• Emotion Recognition - Audio emotion analysis
• Talking Head - Visual speech synthesis

---

Tip: Consider your use case (quality vs speed) and target platform when choosing a TTS model.

This site is open source. Improve this page.