E2 TTS
SWivid/E2-TTS
published Oct 2024 · updated Mar 2025
E2 TTS is a fully non-autoregressive zero-shot text-to-speech model using flow matching.
specs
| Task | Text-to-Speech |
| Architecture | Flat-UNet Transformer |
| License | CC BY-NC 4.0 |
about this model
E2 TTS (Embarrassingly Easy Text-to-Speech) is a fully non-autoregressive, zero-shot text-to-speech model that converts text into speech using a flow-matching-based mel spectrogram generator trained on an audio infilling task. The input text is represented as a character sequence with filler tokens, eliminating the need for additional components such as a duration model, grapheme-to-phoneme conversion, or monotonic alignment search.
The model employs a Flat-UNet Transformer architecture and achieves its zero-shot capability through flow-matching training on large-scale speech data. E2 TTS delivers human-level naturalness and state-of-the-art speaker similarity and intelligibility. According to the original paper, its zero-shot performance is comparable to or surpasses previous systems including Voicebox and NaturalSpeech 3.
E2 TTS is released under the CC BY-NC 4.0 license. On the Hugging Face hub, the model has received over 799,000 monthly downloads and is used in 461 Spaces, with 131 fine-tuned variants and 4 quantizations available. The model is hosted on gigarouter as a managed, OpenAI-compatible API, allowing developers to integrate it directly without managing infrastructure.
For further details, see the paper, the GitHub repository, and demo samples at https://aka.ms/e2tts/.
best for
- ·Zero-shot voice cloning from short audio samples
- ·Generating natural-sounding speech for virtual assistants
FAQ
E2 TTS is a fully non-autoregressive zero-shot text-to-speech model that uses flow matching to generate mel spectrograms from text with filler tokens, achieving human-level naturalness.
The model takes text as input and an audio prompt for voice cloning; generated mel spectrograms are then converted to waveform.
Use the gigarouter OpenAI-compatible endpoint with an API key to send requests for text-to-speech synthesis.
The model is licensed under CC BY-NC 4.0, which allows non-commercial use with attribution.
E2 TTS achieves state-of-the-art speaker similarity and intelligibility comparable to Voicebox and NaturalSpeech 3 while being simpler and fully non-autoregressive.
We're benchmarking and onboarding E2 TTS as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.