README.md

---
language:
  - dv
  - ar
  - en
license: cc-by-nc-4.0
tags:
  - automatic-speech-recognition
  - mms
  - ctc
  - trilingual
  - dhivehi
  - arabic
  - english
datasets:
  - shunyalabs/arabic-speech-dataset
  - shiimi/dhivehi-audio-casts-processed
  - Serialtechlab/dhivehi-mms-v5-combined
  - openslr/librispeech_asr
metrics:
  - wer
base_model: Serialtechlab/mms-trilingual-dv-ar-en
---

# MMS Trilingual ASR v2 - Dhivehi + Arabic + English

Fine-tuned version of mms-trilingual-dv-ar-en with improved:
- **Conversational Arabic** recognition (FLEURS Arabic)
- **Melodic Dhivehi** (Madhaha/podcasts) recognition

## Changes from v1
- Added conversational Arabic data (FLEURS) to replace Quranic-only training
- Added melodic Dhivehi (audio casts) to fix Madhaha confusion with Arabic
- Removed Quranic recitation data

## Training Data
- Arabic: ~2500 samples from FLEURS (conversational)
- Dhivehi Melodic: 1000 samples from audio casts
- Dhivehi Normal: ~1500 samples
- English: ~500 samples from LibriSpeech

## Performance
- Final WER: 0.2820

## Usage
```python
from transformers import AutoProcessor, Wav2Vec2ForCTC
import torch

processor = AutoProcessor.from_pretrained("Serialtechlab/mms-trilingual-dv-ar-en-v2")
model = Wav2Vec2ForCTC.from_pretrained("Serialtechlab/mms-trilingual-dv-ar-en-v2")

# Process audio (16kHz)
inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]
```

## Supported Languages
- Dhivehi (Thaana script) - including melodic/Madhaha
- Arabic (Arabic script) - conversational style
- English (Latin script)