README.md

metadata

license: cc-by-4.0
language:
  - bn
  - hi
  - pa
  - mr
  - gu
base_model:
  - parthiv11/stt_hi_conformer_ctc_large_v2
tags:
  - speech_recognition
  - entity_tagging
  - dialect_prediction
  - gender
  - age
  - intent
library_name: nemo
datasets:
  - WhissleAI/indicvoices_hi_tagged_transcripts
  - WhissleAI/indicvoices_pa_tagged_transcripts
  - WhissleAI/indicvoices_mr_tagged_transcripts

Indo-Aryan Speech Tagger - Conformer CTC Model

This speech tagger performs transcription for 5 Indian Languages: Hindi, Punjabi, Marathi, Bengali and Gujarati. It annotates key entities, predicts speaker age, dialect and intent.

Model Details

• Model Type: NeMo ASR
• Architecture: Conformer CTC
• Language: Bengali, Hindi, Punjabi, Marathi, Gujarati
• Training Data: AI4Bharat IndicVoices Bengali V1 and V2 dataset
• Task: Speech Recognition with Entity Tagging

Usage

import nemo.collections.asr as nemo_asr

# Load model
asr_model = nemo_asr.models.EncDecCTCModel.from_pretrained('WhissleAI/speech-tagger_indo-aryan_ctc_meta')

# Transcribe audio
transcription = asr_model.transcribe(['path/to/audio.wav'])
print(transcription[0])

Model Training

• Base model: Conformer CTC
• Fine-tuned on AI4Bharat IndicVoices Marathi dataset
• Optimized for real-time transcription

License & Attribution

Please cite AI4Bharat when using this model: https://indicvoices.ai4bharat.org/