Update README.md
Browse files
README.md
CHANGED
@@ -1,4 +1,81 @@
|
|
1 |
---
|
2 |
datasets:
|
3 |
- edinburghcstr/ami
|
4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
datasets:
|
3 |
- edinburghcstr/ami
|
4 |
+
base_model:
|
5 |
+
- MIT/ast-finetuned-audioset-10-10-0.4593
|
6 |
+
---
|
7 |
+
# AST-based Speaker Identification on AMI
|
8 |
+
|
9 |
+
## Model description
|
10 |
+
|
11 |
+
This model is a **fine-tuned** version of [MIT/ast-finetuned-audioset-10-10-0.4593](https://huggingface.co/MIT/ast-finetuned-audioset-10-10-0.4593)
|
12 |
+
for speaker classification on the AMI Meeting Corpus. It was trained on **50** speakers (adjust `num_labels` if different), using 128-bin mel-spectrograms of 1024 frames.
|
13 |
+
|
14 |
+
- **Base architecture**: Audio Spectrogram Transformer (AST)
|
15 |
+
- **Training**: ~10 epochs, batch size=4, learning rate=1e-5, AdamW optimizer, mixed precision
|
16 |
+
- **Data**: Stratified samples from AMI train/validation/test splits
|
17 |
+
- **Performance**: Not good, this was just a small experiment for diarization
|
18 |
+
|
19 |
+
|
20 |
+
|
21 |
+
## How to use
|
22 |
+
|
23 |
+
```python
|
24 |
+
from transformers import AutoProcessor, ASTForAudioClassification
|
25 |
+
import torch
|
26 |
+
import numpy as np
|
27 |
+
|
28 |
+
# 1) Load the model and processor
|
29 |
+
MODEL_ID = "agutig/AST_diarizer"
|
30 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID)
|
31 |
+
model = ASTForAudioClassification.from_pretrained(MODEL_ID)
|
32 |
+
model.eval()
|
33 |
+
|
34 |
+
# 2) Prepare a 1-second audio sample (or load your own)
|
35 |
+
sr = 16000
|
36 |
+
audio = np.random.randn(sr).astype(np.float32)
|
37 |
+
# Alternatively:
|
38 |
+
# import librosa
|
39 |
+
# audio, _ = librosa.load("your_audio.wav", sr=sr)
|
40 |
+
|
41 |
+
# 3) Preprocess and run inference
|
42 |
+
inputs = processor(audio, sampling_rate=sr, return_tensors="pt")
|
43 |
+
with torch.no_grad():
|
44 |
+
logits = model(**inputs).logits # shape [1, num_labels]
|
45 |
+
probs = torch.softmax(logits, dim=-1)[0]
|
46 |
+
pred_i = int(probs.argmax())
|
47 |
+
|
48 |
+
print(f"Predicted speaker index: {pred_i}")
|
49 |
+
```
|
50 |
+
|
51 |
+
## Usage with `pipeline`
|
52 |
+
|
53 |
+
```python
|
54 |
+
from transformers import pipeline
|
55 |
+
|
56 |
+
speaker_id = pipeline(
|
57 |
+
task="audio-classification",
|
58 |
+
model="agutig/AST_diarizer",
|
59 |
+
return_all_scores=True
|
60 |
+
)
|
61 |
+
|
62 |
+
results = speaker_id("path/to/audio.wav")
|
63 |
+
print(results)
|
64 |
+
```
|
65 |
+
|
66 |
+
## Evaluation & Benchmarks
|
67 |
+
|
68 |
+
| Metric | Value |
|
69 |
+
|--------------------------|---------|
|
70 |
+
| Accuracy (test) | 0.XX |
|
71 |
+
| Adjusted Rand Index (ARI)| 0.YY |
|
72 |
+
| Normalized Mutual Info | 0.ZZ |
|
73 |
+
|
74 |
+
_Fill in actual values._
|
75 |
+
|
76 |
+
|
77 |
+
|
78 |
+
## License
|
79 |
+
|
80 |
+
- **Model**: Apache 2.0
|
81 |
+
- **Base code (AST AudioSet)**: MIT License
|