language: sw
thumbnail: null
pipeline_tag: automatic-speech-recognition
tags:
- CTC
- pytorch
- speechbrain
- Transformer
license: apache-2.0
datasets:
- commonvoice
metrics:
- wer
- cer
wav2vec 2.0 with CTC/Attention trained on DVoice Swahili (No LM)
This repository provides all the necessary tools to perform automatic speech recognition from an end-to-end system pretrained on a DVoice-VoxLingua107 Swahili dataset within SpeechBrain. For a better experience, we encourage you to learn more about SpeechBrain.
| DVoice Release | Val. CER | Val. WER | Test CER | Test WER |
|---|---|---|---|---|
| v2.0 | 8.83 | 22.78 | 9.46 | 23.16 |
About DVoice
DVoice is a community initiative that aims to provide African languages and dialects with data and models to facilitate their use of voice technologies. The lack of data on these languages makes it necessary to collect data using methods that are specific to each language. Two different approaches are currently used: the DVoice platforms (https://dvoice.ma and https://dvoice.sn), which are based on Mozilla Common Voice, for collecting authentic recordings from the community, and transfer learning techniques for automatically labeling the recordings. The DVoice platform currently manages 7 languages including Darija (Moroccan Arabic dialect) whose dataset appears on this version, Wolof, Mandingo, Serere, Pular, Diola and Soninke.
This Swahili ASR model was obtained by combining the VoxLingua107 Swahili Dataset with transfer learning and speech augmentation modules.
Pipeline description
This ASR system is composed of 2 different but linked blocks:
- Tokenizer (unigram) that transforms words into subword units and trained with the train transcriptions.
- Acoustic model (wav2vec2.0 + CTC). A pretrained wav2vec 2.0 model (facebook/wav2vec2-large-xlsr-53) is combined with two DNN layers and finetuned on the Darija dataset. The obtained final acoustic representation is given to the CTC greedy decoder. The system is trained with recordings sampled at 16kHz (single channel). The code will automatically normalize your audio (i.e., resampling + mono channel selection) when calling transcribe_file if needed.
Install SpeechBrain
First of all, please install tranformers and SpeechBrain with the following command:
pip install speechbrain transformers
Please notice that we encourage you to read the SpeechBrain tutorials and learn more about SpeechBrain.
Transcribing your own audio files (in Swahili)
from speechbrain.pretrained import EncoderASR
asr_model = EncoderASR.from_hparams(source="nairaxo/dvoice-swahili", savedir="pretrained_models/asr-wav2vec2-dvoice-sw")
asr_model.transcribe_file('./the_path_to_your_audio_file')
Inference on GPU
To perform inference on the GPU, add run_opts={"device":"cuda"} when calling the from_hparams method.
Training
To train the model from scratch, please see our GitHub tutorial here.
Limitations
The SpeechBrain team does not provide any warranty on the performance achieved by this model when used on other datasets.
Referencing SpeechBrain
@misc{SB2021,
author = {Ravanelli, Mirco and Parcollet, Titouan and Rouhe, Aku and Plantinga, Peter and Rastorgueva, Elena and Lugosch, Loren and Dawalatabad, Nauman and Ju-Chieh, Chou and Heba, Abdel and Grondin, Francois and Aris, William and Liao, Chien-Feng and Cornell, Samuele and Yeh, Sung-Lin and Na, Hwidong and Gao, Yan and Fu, Szu-Wei and Subakan, Cem and De Mori, Renato and Bengio, Yoshua },
title = {SpeechBrain},
year = {2021},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\\\\url{https://github.com/speechbrain/speechbrain}},
}
About SpeechBrain
SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to be simple, extremely flexible, and user-friendly. Competitive or state-of-the-art performance is obtained in various domains. Website: https://speechbrain.github.io/ GitHub: https://github.com/speechbrain/speechbrain