Add model card
Browse files
README.md
CHANGED
|
@@ -1,5 +1,63 @@
|
|
| 1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
license: other
|
| 3 |
-
|
| 4 |
-
|
|
|
|
|
|
|
|
|
|
| 5 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
language:
|
| 3 |
+
- hi
|
| 4 |
+
library_name: transformers
|
| 5 |
+
pipeline_tag: automatic-speech-recognition
|
| 6 |
license: other
|
| 7 |
+
tags:
|
| 8 |
+
- whisper
|
| 9 |
+
- speech-recognition
|
| 10 |
+
- hindi
|
| 11 |
+
- hinglish
|
| 12 |
---
|
| 13 |
+
|
| 14 |
+
# MME_STT_Model
|
| 15 |
+
|
| 16 |
+
Hindi speech to Roman-script Hinglish transcription. This model was fine-tuned with LoRA and the adapter
|
| 17 |
+
was merged into the weights, so it loads as a normal Transformers model with no PEFT.
|
| 18 |
+
|
| 19 |
+
## Usage
|
| 20 |
+
|
| 21 |
+
```python
|
| 22 |
+
import torch, librosa
|
| 23 |
+
from transformers import WhisperForConditionalGeneration, WhisperProcessor
|
| 24 |
+
|
| 25 |
+
model = WhisperForConditionalGeneration.from_pretrained("MmeOG/MME_STT_Model", torch_dtype=torch.float16).to("cuda")
|
| 26 |
+
processor = WhisperProcessor.from_pretrained("MmeOG/MME_STT_Model")
|
| 27 |
+
|
| 28 |
+
audio, _ = librosa.load("sample.wav", sr=16000)
|
| 29 |
+
feats = processor(audio, sampling_rate=16000, return_tensors="pt").input_features.to("cuda", torch.float16)
|
| 30 |
+
ids = model.generate(feats, language="hi", task="transcribe")
|
| 31 |
+
print(processor.batch_decode(ids, skip_special_tokens=True)[0])
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
## Training
|
| 35 |
+
|
| 36 |
+
LoRA fine-tune on about 190k utterances: IndicVoices (about 92%), MUCS and UJS (about 4% each).
|
| 37 |
+
Two epochs, batch size 16, learning rate 3e-5 (cosine), one A100 80GB.
|
| 38 |
+
|
| 39 |
+
## Evaluation
|
| 40 |
+
|
| 41 |
+
Phonetic WER (pWER, lower is better) on the public set `RinggAI/ASR-Benchmarking-Dataset`
|
| 42 |
+
(sampled 200 clips per dataset, 1,100 clips in total). The merged model reproduces the adapter's published scores to within 0.05 points.
|
| 43 |
+
|
| 44 |
+
| Dataset | This model | Before fine-tuning) |
|
| 45 |
+
|---|---|---|
|
| 46 |
+
| fleurs | 18.45 | 20.15 |
|
| 47 |
+
| commonvoice | 26.42 | 29.16 |
|
| 48 |
+
| indictts | 13.32 | 13.72 |
|
| 49 |
+
| kathbath | 19.79 | 21.73 |
|
| 50 |
+
| kathbath_noisy | 21.91 | 23.37 |
|
| 51 |
+
| mucs | 19.65 | 20.84 |
|
| 52 |
+
| **all** | **19.64** | **21.21** |
|
| 53 |
+
|
| 54 |
+
pWER is a lenient metric that compares text in a shared phonetic form, so it ranks systems on the
|
| 55 |
+
same clips but should not be read as an absolute accuracy figure.
|
| 56 |
+
|
| 57 |
+
## Limitations
|
| 58 |
+
|
| 59 |
+
- Not a streaming model: it transcribes after the speaker finishes.
|
| 60 |
+
- Short utterances (under 2 seconds) have high error rates.
|
| 61 |
+
- Accuracy drops sharply in loud background noise, especially babble.
|
| 62 |
+
- Not evaluated on real phone-call recordings.
|
| 63 |
+
- Commercial systems scored better on the same public benchmark.
|