fd-speech-demo / README.md
voidful's picture
Rebrand demo as FDSpeech
66053d0 verified
|
Raw
History Blame Contribute Delete
1.2 kB
---
title: FDSpeech TTS Comparison
emoji: 🎧
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.9.0
app_file: app.py
python_version: "3.10"
pinned: false
license: apache-2.0
short_description: Compare VoxCPM2 4/10-step and FDSpeech 4-step speech
models:
- voidful/FDSpeech-VoxCPM2
---
# FDSpeech: four-step VoxCPM2 comparison
This Space provides aligned Seed-TTS English listening examples for:
- original VoxCPM2 at four steps;
- original VoxCPM2 at ten steps; and
- **FDSpeech** at four steps.
FDSpeech is the released VoxCPM2 LoRA adapter trained with a
Fréchet-distance loss on speech representations. It improves upstream
Seed-TTS English WER from 2.2279% for the original four-step sampler to
1.4147%, without adding an inference-time module.
- Model: [`voidful/FDSpeech-VoxCPM2`](https://huggingface.co/voidful/FDSpeech-VoxCPM2)
- Code: [`voidful/fd-speech`](https://github.com/voidful/fd-speech)
- Paper: [Fréchet Distance Loss on Speech Representations for Text-to-Speech Synthesis](https://arxiv.org/abs/2607.06027)
The Space serves checked-in, aligned audio so every system uses the same prompt
and evaluation seed. Negative cases are included alongside improvements.