text → phonemes → mel → waveform
text frontend
normalize + G2P
"6 pm" → "six p m"
acoustic model
phones → mel [T, 80]
FastSpeech / VITS / F5
vocoder
mel → wav [T·256]
HiFi-GAN, BigVGAN
waveform 24 kHz
float32 in [-1, 1]
speaker.wav
open-source TTS timeline
Tacotron 2
2017 · AR seq2seq
FastSpeech 2
2020 · non-AR
VITS
2021 · end-to-end flow
XTTS v2
2024 · voice cloning
F5-TTS
2024 · flow matching
Kokoro
2024 · 82M CPU
2026 open-source leaderboard (LibriTTS test-clean)
model
UTMOS
CER via Whisper
params
ground truth 4.08 1.2% —
F5-TTS 3.95 2.1% 335M
Kokoro v0.19 3.87 1.8% 82M
XTTS v2 3.81 3.5% 470M