gradient-queen/whisper-tiny-indian-digits-final

Whisper-Tiny Indian English Digit Recognizer 🎙️

This model is a fine-tuned version of openai/whisper-tiny trained on the gradient-queen/spoken-digits-indian-english dataset. It is specialized in transcribing spoken digits (from 1 to 10) spoken in Indian English accents.

Model Description

• Developed by: gradient-queen
• Model Type: Sequence-to-Sequence Audio Transformer (Whisper)
• Language(s): English (specifically Indian English accent variations)
• License: MIT
• Finetuned from model: openai/whisper-tiny

Intended Uses & Limitations

This model is intended to recognize verbal numbers from 1 to 10 (represented textually as word labels like "one", "two", "three", etc.) in Indian English accents.

Limitations

• Optimized specifically for short, isolated digit pronunciations.
• Performance might degrade on conversational sentences, continuous digits, or non-English audio.

Training and Evaluation Data

• Dataset Name: gradient-queen/spoken-digits-indian-english
• Audio Format: Resampled to 16,000 Hz, single-channel PCM.
• Target Text: Fully normalized lowercase English word transcriptions ("zero", "one", ..., "ten").

How to Use

You can use this model directly via Hugging Face pipelines for inference:

import torch
from transformers import pipeline

# Initialize pipeline
device = "cuda:0" if torch.cuda.is_available() else "cpu"
pipe = pipeline(
    "automatic-speech-recognition",
    model="gradient-queen/whisper-tiny-indian-digits-final",
    device=device
)

# Run prediction
result = pipe("path_to_audio_file.wav")
print("Transcription:", result["text"])

Training Hyperparameters

• Learning Rate: 1e-4
• Batch Size: 16
• Optimizer: AdamW
• Warmup Steps: 50
• Max Steps: 500
• Precision: Mixed Precision (fp16)