File size: 2,613 Bytes
4926b2d faf6ecb 54e0da1 faf6ecb 54e0da1 17044dc faf6ecb 4926b2d 52a2eb8 4926b2d 3dc5507 faf6ecb 3dc5507 4926b2d faf6ecb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 |
---
license: mit
task_categories:
- automatic-speech-recognition
tags:
- collator
---
## Dynamic Audio Data Augmentation
## Key Benefits
Enhanced Robustness: By varying spectrogram parameters and injecting realistic noise, our models learn to handle a wide range of audio conditions.
Low Overhead: The augmentation is integrated into the existing pipeline, ensuring minimal additional computational cost. Data collator (low overhead) versus Dataset (higher overhead)
### On-the-Fly Spectrogram Parameter Adjustment:
n_fft and hop_length: Values for n_fft and hop_length are randomly selected from predefined ranges for each audio sample, providing varied spectrogram representations.
### Log-Mel Modulation:
Augmentation process integrates with the existing log-Mel spectrogram calculation. This means we modulate the parameters of the log-Mel spectrogram dynamically, ensuring no additional overhead is introduced while providing effective data augmentation.
### Efficiency and Performance
Log-Mel Spectrogram Manipulation:
Augmentation process seamlessly integrates into the existing log-Mel spectrogram calculation, adding no extra overhead. This efficient design ensures that our preprocessing remains computationally lightweight and fast.
#### Adaptive Context-Aware Noise Injection
Preprocessing pipeline that includes adaptive context-aware noise injection to enhance model robustness. This method dynamically adjusts noise intensity based on the amplitude of the audio signal, ensuring realistic and effective augmentation.
- **Types of Noise**: White, pink, and environmental noise.
- **Dynamic Adjustment**: Noise intensity is scaled based on the amplitude of the audio signal.
- **Integration**: The noise injection process is seamlessly integrated into our existing log-Mel spectrogram calculation pipeline, adding minimal overhead.
##### Key Benefits
- **Improved Generalization**: Models become more resilient to noise and diverse audio conditions.
- **Low Overhead**: The augmentation process leverages the existing pipeline, ensuring efficient computation without significant additional cost.
##### Example Usage
```python
## HF transformers or pure pytorch
data_collator = DataCollatorSpeechSeq2SeqWithPadding(
processor=processor,
decoder_start_token_id=model.config.decoder_start_token_id,
apply_augmentation=True,
apply_noise_injection=True # Enable adaptive noise injection
)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=2, shuffle=True, collate_fn=data_collator)
for batch in dataloader:
outputs = model(batch)
|