abhishekgautamm/whisper-tiny-hinglish-ct2

Whisper-tiny Hinglish — CTranslate2 (romanized)

A CTranslate2 export of a fine-tuned openai/whisper-tiny that transcribes Hindi / English / Hinglish speech directly into romanized text (kya scene hai, not Devanagari). Built for fast, CPU-only, on-device use with faster-whisper.

This repo is the full-precision (float32) build. An int8 build (~46 MB) ships inside the Yapper app. The PyTorch/Transformers model card is at ABHISHEKgauti25/whisper-tiny-hinglish.

Usage

from faster_whisper import WhisperModel

model = WhisperModel("ABHISHEKgauti25/whisper-tiny-hinglish-ct2", device="cpu", compute_type="float32")
segments, _ = model.transcribe("clip.wav", language="en", beam_size=1)
print("".join(s.text for s in segments).strip())

Set compute_type="int8" to quantize on load for lower memory/latency.

Notes

• Output is romanized Hindi/English/Hinglish; English spans stay verbatim.
• Small model — fast and convenient, not state-of-the-art. Expect errors on noisy audio, fast speech, or uncommon words; romanized spelling is phonetic and can vary.
• License: base model is MIT; training datasets (HiACC, MUCS, Common Voice) carry their own terms — verify before commercial use. apache-2.0 above is a placeholder to adjust.

Full training details and limitations: see the base model card ABHISHEKgauti25/whisper-tiny-hinglish.