E2 TTS

E2 TTS (Embarrassingly Easy Text-to-Speech) is a zero-shot speech synthesis model published by Microsoft researchers in June 2024 and accepted to IEEE SLT 2024. It is fully non-autoregressive: text becomes a character sequence with filler tokens, and a flow-matching mel spectrogram generator trained on audio infilling produces the speech.

The design needs no duration model, grapheme-to-phoneme converter, or monotonic alignment search, yet matches or beats systems like Voicebox and NaturalSpeech 3 on speaker similarity and intelligibility with human-level naturalness.

The paper proposes inference-friendly variants for flexible input representations. No official weights were released, but the simplicity of the approach made it influential in the flow-matching TTS wave.