| Property | Value |
|---|---|
| License | CC-BY-NC-4.0 |
| Framework | Fairseq |
| Paper | Research Paper |
| Task | Speech-to-Speech Translation |
What is unit_hifigan_mhubert_vp_en_es_fr_it3_400k_layer11_km1000_lj_dur?
This is a sophisticated speech-to-speech translation model developed by Facebook, specifically designed for Spanish-to-English translation. The model leverages the HiFi-GAN vocoder architecture combined with MHuBERT features, trained on a diverse set of datasets including mTEDx, CoVoST 2, Europarl-ST, and VoxPopuli.
Implementation Details
The model implements a direct speech-to-speech translation approach using discrete units, as detailed in Facebook's S2UT framework. It utilizes layer 11 features with 1000 k-means clusters and incorporates duration modeling for improved speech synthesis.
- Implements HiFi-GAN vocoder architecture
- Utilizes MHuBERT features from layer 11
- Trained on 400k steps
- Incorporates duration modeling for natural speech output
Core Capabilities
- Direct Spanish to English speech translation
- High-quality speech synthesis using HiFi-GAN
- Support for 16000Hz mono channel audio input
- Integration with Fairseq's hub interface for easy deployment
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its direct speech-to-speech translation approach without intermediate text representation, utilizing discrete units and advanced vocoder technology for high-quality output.
Q: What are the recommended use cases?
The model is ideal for applications requiring Spanish-to-English speech translation, particularly in scenarios where maintaining natural speech qualities is important, such as conference interpretation or multimedia content translation.