autotinglishsub-whisper-telugu

AutoTinglishSub – Whisper Telugu Tinglish Fine-Tune

Overview

cvas-544/autotinglishsub-whisper-telugu is a fine-tuned Automatic Speech Recognition (ASR) model optimized for Tinglish — Telugu and Indian English code-mixed speech.

This model is designed specifically for subtitle generation workflows in short-form content such as Reels, TikTok, and YouTube Shorts.

It improves transcription quality for:

• Telugu conversational speech
• Indian English accents
• Telugu-English code-switching
• Fast-paced creator-style delivery
• Informal spoken content

---

Model Lineage

This model is fine-tuned from:

• vasista22/whisper-telugu-large-v2 (Apache 2.0)

Which itself is fine-tuned from:

• openai/whisper-large-v2 (MIT License)

This model continues that lineage with further optimization for Tinglish subtitle workflows.

---

Intended Use

Primary use cases:

• Subtitle generation for short-form video
• Creator-focused ASR pipelines
• Local CLI subtitle tools
• Tinglish speech recognition
• Code-switched Telugu + English audio

The model is optimized for conversational and creator-style speech rather than formal Telugu literature.

---

Training Summary

• Fine-tuning performed on Tinglish-focused dataset
• Total training chunks used: 120
• Objective: Improve transcription accuracy for Telugu + Indian English mixed speech
• Focus on subtitle-ready clean outputs

---

Evaluation Metrics

Word Error Rate (WER)

Details:

• Measured after first fine-tuning phase
• Evaluated on held-out validation split
• Based on 120 training chunks
• Metric computed using standard WER calculation (substitutions + insertions + deletions / total words)

This represents a significant improvement in handling Tinglish code-mixed speech compared to the base Telugu fine-tuned model in informal creator-style audio.

---

Usage

Using Transformers Pipeline

import torch
from transformers import pipeline

audio = "/path/to/audio.wav"
device = "cuda:0" if torch.cuda.is_available() else "cpu"

transcribe = pipeline(
    task="automatic-speech-recognition",
    model="cvas-544/autotinglishsub-whisper-telugu",
    chunk_length_s=30,
    device=device
)

print("Transcription:", transcribe(audio)["text"])