Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: it
|
| 3 |
+
license: mit
|
| 4 |
+
tags:
|
| 5 |
+
- whisper
|
| 6 |
+
- automatic-speech-recognition
|
| 7 |
+
- italian
|
| 8 |
+
- ctranslate2
|
| 9 |
+
- faster-whisper
|
| 10 |
+
- whisperx
|
| 11 |
+
- localai
|
| 12 |
+
datasets:
|
| 13 |
+
- mozilla-foundation/common_voice_25_0
|
| 14 |
+
base_model: openai/whisper-tiny
|
| 15 |
+
pipeline_tag: automatic-speech-recognition
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# whisper-tiny-it
|
| 19 |
+
|
| 20 |
+
Fine-tuned [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) (39M params) for Italian automatic speech recognition (ASR).
|
| 21 |
+
|
| 22 |
+
**Author:** Ettore Di Giacinto
|
| 23 |
+
|
| 24 |
+
Brought to you by the [LocalAI](https://github.com/mudler/LocalAI) team. This model can be used directly with [LocalAI](https://localai.io).
|
| 25 |
+
|
| 26 |
+
## Results
|
| 27 |
+
|
| 28 |
+
Evaluated on Common Voice 25.0 Italian test set (15,184 samples):
|
| 29 |
+
|
| 30 |
+
| Step | Train Loss | Eval Loss | WER |
|
| 31 |
+
|------|-----------|-----------|-----|
|
| 32 |
+
| 1000 | — | 0.59 | 37.1% |
|
| 33 |
+
| 3000 | 0.42 | 0.47 | 30.8% |
|
| 34 |
+
| 5000 | — | 0.43 | 28.7% |
|
| 35 |
+
| 10000 | 0.29 | 0.40 | **27.1%** |
|
| 36 |
+
|
| 37 |
+
## Training Details
|
| 38 |
+
|
| 39 |
+
- **Base model:** openai/whisper-tiny (39M parameters)
|
| 40 |
+
- **Dataset:** Common Voice 25.0 Italian (173k train, 15k dev, 15k test)
|
| 41 |
+
- **Steps:** 10,000 (batch size 32, ~1.8 epochs)
|
| 42 |
+
- **Learning rate:** 1e-5 with 500 warmup steps
|
| 43 |
+
- **Precision:** bf16 on NVIDIA GB10
|
| 44 |
+
- **Training time:** ~2 hours
|
| 45 |
+
|
| 46 |
+
## Usage
|
| 47 |
+
|
| 48 |
+
### Transformers
|
| 49 |
+
|
| 50 |
+
```python
|
| 51 |
+
from transformers import pipeline
|
| 52 |
+
|
| 53 |
+
pipe = pipeline("automatic-speech-recognition", model="LocalAI-io/whisper-tiny-it")
|
| 54 |
+
result = pipe("audio.mp3", generate_kwargs={"language": "it", "task": "transcribe"})
|
| 55 |
+
print(result["text"])
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
### CTranslate2 / faster-whisper
|
| 59 |
+
|
| 60 |
+
For optimized CPU inference, use the INT8 quantized version: [LocalAI-io/whisper-tiny-it-ct2-int8](https://huggingface.co/LocalAI-io/whisper-tiny-it-ct2-int8) (39MB).
|
| 61 |
+
|
| 62 |
+
### LocalAI
|
| 63 |
+
|
| 64 |
+
This model is compatible with [LocalAI](https://github.com/mudler/LocalAI) for local, self-hosted AI inference.
|
| 65 |
+
|
| 66 |
+
## Links
|
| 67 |
+
|
| 68 |
+
- **Code:** [github.com/localai-org/whisper-it](https://github.com/localai-org/whisper-it)
|
| 69 |
+
- **CTranslate2 INT8:** [LocalAI-io/whisper-tiny-it-ct2-int8](https://huggingface.co/LocalAI-io/whisper-tiny-it-ct2-int8)
|
| 70 |
+
- **LocalAI:** [github.com/mudler/LocalAI](https://github.com/mudler/LocalAI)
|