Serialtechlab / mms-trilingual-dv-ar-en-v3-clean

MMS Trilingual ASR v3 - Dhivehi + Arabic + English (Madhaha Fix)

Fine-tuned version of mms-trilingual-dv-ar-en-v2 with fixed Madhaha recognition .

Problem Solved

v2 model confused melodic Dhivehi (Madhaha/religious songs) with Arabic, outputting Arabic script instead of Thaana. This version fixes that issue.

Training Strategy

• Started from v2 model (preserves improved English/Arabic recognition)
• Trained ONLY on Dhivehi data (no Arabic interference)
• Oversampled melodic Dhivehi 3x to emphasize the pattern
• Higher learning rate (3e-05) to change associations aggressively
• 5 epochs for stronger reinforcement

Training Data

• Melodic Dhivehi: ~3000 samples (oversampled from audio casts)
• Normal Dhivehi: ~1500 samples

Performance

• Final WER: 0.2153

Usage

from transformers import AutoProcessor, Wav2Vec2ForCTC
import torch

processor = AutoProcessor.from_pretrained("Serialtechlab/mms-trilingual-dv-ar-en-v3")
model = Wav2Vec2ForCTC.from_pretrained("Serialtechlab/mms-trilingual-dv-ar-en-v3")

# 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) - preserved from v2
• English (Latin script) - preserved from v2 (improved Thaana transliteration)

Changes from v2

• v3 specifically targets the Madhaha confusion issue
• Melodic Dhivehi now correctly outputs Thaana script
• Preserves v2's improved English and Arabic recognition