Model Card for Finetuned Version of Whisper-Medium
This model was trained on a subset of the synthetically generated data that later on was filtered to increase the performance of Whisper Model. The approach involves aligning representations of synthetic audio and corresponding text transcripts to identify and remove low-quality samples, improving the overall training data quality.
In this Specific Model we used 96,08% of synthetic data generated by SeamllesMT4LargeV2, the rest was removed by the filtering model. The training set also contained, the CommonVoice Dataset, Multilibri Speach, and Bracarense (Fully Portuguese Dialect)
Model Details
- Developed by: Yuriy Perezhohin, Tiago Santos, Victor Costa, Fernando Peres, and Mauro Castelli.
- Funded by: MyNorth AI Research
- Shared by: MyNorth AI Research
- Model type: ASR with contrastive learning-based synthetic data filtering
- Language: Portuguese
- License: APACHE 2.0
- Finetuned from model: Whisper Small
Model Sources
- Repository: https://github.com/my-north-ai/semantic_audio_filtering
- Paper: Comming Soon
Uses
This model can be directly used for improving ASR systems in Portuguese, particularly in scenarios with limited real-world data or unique linguistic characteristics.
Out-of-Scope Use
The model is not suitable for tasks involving languages other than Portuguese without additional fine-tuning and data adjustments.
Bias, Risks, and Limitations
Users should be aware of potential biases introduced by synthetic data and ensure the quality of the data aligns with the target application's requirements. It is recommended to evaluate the model's performance on diverse datasets to identify and mitigate biases.
How to Get Started with the Model
from transformers import pipeline
model = pipeline("automatic-speech-recognition", model="my-north-ai/whisper-medium-pt")
result = model("path_to_audio_file.wav")
print(result)
Training Details
Training Data
The training data includes 140 hours of synthetically generated Portuguese speech and transcripts, along with real data from the Multilingual LibriSpeech Corpus (MLS), Common Voice (CV) 16.1, and the Perfil Sociolinguístico da Fala Bracarense (PSFB) dataset
Training Procedure
The model was fine tuned using DDP methodolgy across 4 A10g GPUS
Preprocessin
The preprocessing steps include text normalization, removal of special characters, and ensuring consistent formatting for TTS generation.
Training Hyperparameters
- Training regime: fp16 mixed precision
- Learning Rate: 1e-5
- Batch Size: 32
- Epochs 3
Evaluation
Testing Data, Factors & Metrics
Testing Data
The testing data includes subsets from the FLEURS dataset and PSFB, chosen for their linguistic diversity and unique speech patterns.
Metrics
Evaluation Results
Word Error Rate (WER) Comparison
WER results for FLEURS for the fine-tuned model versus pretrained model with and without text normalization.
| Model Size | Model Type | WER (Normalized) | WER (Non-Normalized) |
|---|---|---|---|
| Small | Pretrained | 10.87 | 15.43 |
| Small | FS-17.68% | 10.45 | 18.57 |
| Small | FS-3.92% | 10.34 | 18.53 |
| Small | FS-0.24% | 10.58 | 18.90 |
| Small | Zero Synthetic | 10.90 | 19.32 |
| Medium | Pretrained | 8.62 | 12.65 |
| Medium | FS-17.68% | 6.58 | 14.46 |
| Medium | FS-3.92% | 6.57 | 14.44 |
| Medium | FS-0.24% | 6.58 | 14.54 |
| Medium | Zero Synthetic | 6.97 | 14.74 |
| Large V3 | Pretrained | 7.70 | 11.78 |
| Large V3 | FS-17.68% | 4.73 | 10.83 |
| Large V3 | FS-3.92% | 4.65 | 11.09 |
| Large V3 | FS-0.24% | 4.80 | 11.28 |
| Large V3 | Zero Synthetic | 4.86 | 10.92 |
Environmental Impact
Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
- Hardware Type: NVIDIA A10G
- Hours used: 15
- Cloud Provider: AWS
- Compute Region: US EAST
- Downloads last month
- 18