Whisper Large v3 Turbo — Hindi Fine-Tuned
Fine-tuned openai/whisper-large-v3-turbo on Monsoon Hindi ASR data using LoRA (r=32, alpha=64).
Results
Open ASR Leaderboard (official, evaluated on HF H200 GPUs)
| Benchmark | WER / OIWER | RTFx |
|---|---|---|
| Monsoon hi-IN (OIWER) | 22.82% | 275.69 |
| AMI-Cleaned | 12.79% | 436.75 |
| Earnings22 | 7.37% | 1426.70 |
| GigaSpeech | 8.10% | 731.01 |
| LibriSpeech Clean | 1.68% | 724.74 |
| LibriSpeech Other | 4.02% | 670.05 |
| SPGISpeech | 2.76% | 964.77 |
| Monsoon en-IN | 3.86% | 975.66 |
| VoxPopuli | 3.17% | 864.82 |
| English Mean WER | 5.47% | 838.65 |
For reference, the base model openai/whisper-large-v3-turbo scores 37.98% OIWER on Monsoon hi-IN — a 40% relative improvement.
Anti-forgetting benchmarks (local evaluation)
| Benchmark | Baseline | Fine-tuned | Delta |
|---|---|---|---|
| LibriSpeech Clean (WER) | 3.70% | 3.94% | +0.24pp |
| FLEURS English (WER) | 4.86% | 5.31% | +0.45pp |
| FLEURS Hindi (WER) | 32.58% | 34.94% | +2.36pp |
Training Details
- Method: LoRA (Low-Rank Adaptation)
- Rank: 32, Alpha: 64, Dropout: 0.05
- Target modules: q_proj, k_proj, v_proj, out_proj, fc1, fc2
- Trainable parameters: 27.8M / 809M (3.33%)
- Learning rate: 1e-4 with cosine schedule, 500 warmup steps
- Epochs: 3 (729 total steps)
- Training data: ~5.8 hours (Monsoon hi-IN train split)
- GPU: NVIDIA A10G (24GB)
- Training time: ~2.5 hours
- Total cost: ~$15
Usage
from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch
model = WhisperForConditionalGeneration.from_pretrained(
"annotehrushi/whisper-large-v3-turbo-hindi",
torch_dtype=torch.float16,
)
processor = WhisperProcessor.from_pretrained(
"annotehrushi/whisper-large-v3-turbo-hindi"
)
model.generation_config.language = "hindi"
model.generation_config.task = "transcribe"
Open ASR Leaderboard
Submission: PR #219