Downloads · 30 days
116
1% of all-time downloads
caiocrocha/wav2vec2-large-xlsr-53-phoneme-portuguese
wav2vec2-large-xlsr-53-phoneme-portuguese is a automatic speech recognition model from caiocrocha. Use it when you need speech turned into text. It is set up for transformers. The card lists the license as apache-2.0.
Link to thesis in the academic repository: https://repositorio.ufjf.br/jspui/handle/ufjf/19250. Developed for a Computational Engineering undergraduate thesis at the Federal University of Juiz de Fora (UFJF), defended…
Downloads · 30 days
116
1% of all-time downloads
All-time downloads
12.6K
Public
Parameters
315M
2.5 GB on disk
Likes
1
Public
Click a slice to open those files.
.bin1.3 GB · 50%
From the Hugging Face model README
Potential factors that might affect the model's performance are age, accent, speech rate, fluency, speech idiosyncrasies, background noise, reverberation, and transmission channel phenomena.
The model was evaluated on the sample test and dev sets, obtaining the following results:
| Set | PER | Accuracy |
|---|---|---|
| Dev | 0.2197 | 0.89 |
| Test | 0.1587 | 0.16 |
Additionaly, we evaluated its performance across the phone classes, presented below:


The model is biased towards the accent featured in FalaBrasil's G2P tool, which was used to transcribe the CORAA ASR datasets. Furthermore, even though the corpus contains several PT-BR accents (Recife, Minas Gerais, standard and non-standard São Paulo accents, among others), the model may underperform on speech featuring underrepresented accents.