Credits
SmollerTTS pairs our acoustic model with the pretrained HiFi-GAN LJ_V2 vocoder. These are the papers, data and tools behind it.
Papers & data
- HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis
Jungil Kong, Jaehyeon Kim and Jaekyoung Bae. NeurIPS, 2020.
Our waveform generator uses the published LJ_V2 weights.
- FastSpeech 2: Fast and High-Quality End-to-End Text to Speech
Yi Ren et al. ICLR, 2021.
Architectural reference for duration, pitch and energy conditioning.
- Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi
Michael McAuliffe et al. Interspeech, 2017.
Alignment tools used to prepare training targets.
- UTMOS: UTokyo-SaruLab System for VoiceMOS Challenge 2022
Takaaki Saeki et al. Interspeech, 2022.
The automatic quality estimator used in the comparison chart.
- The LJ Speech Dataset
The English recordings used to train our acoustic model.
Software & sources
ONNX Runtime runs inference in the browser. Pronunciation uses MFA dictionaries and phonemize; number reading uses to-words.
Published comparison results come from tts-bench. Our separate 48-sentence test compares SmollerTTS with Tiny Voca, using UTMOS and faster-whisper. The interface uses the Geist typeface.