Whisper-Small Hindi Fine-Tuned Model

Overview

This is a fine-tuned version of OpenAI Whisper-small model for Hindi Automatic Speech Recognition (ASR).
The model is trained on ~10 hours of Hindi conversational speech data containing real-world noise, disfluencies, and natural speaking patterns.

---

Dataset

• ~10 hours Hindi conversational dataset
• Includes real-world speech with fillers, pauses, and variations
• Audio resampled to 16kHz mono
• Transcriptions processed from segment-level JSON files

---

Training Details

• Base model: Whisper-small
• Framework: Hugging Face Transformers
• Training environment: Google Colab (GPU)
• Epochs: 2
• Mixed precision training (fp16)

---

Results

👉 WER Reduction: ~61.6%

Additionally, on clean Hindi speech samples, the model achieves approximately ~0.15 WER, indicating strong performance in low-noise conditions.

---

Evaluation

• Dataset: FLEURS Hindi (hi_in)
• Hindi language enforced during decoding
• Text normalization applied before WER computation
• Evaluation performed on 418 samples

---

Usage

from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="Satyam0077/whisper-small-hindi-finetuned-7tym",
    generate_kwargs={"language": "hi"}
)

result = pipe("audio.wav")
print(result["text"])