You can deploy this model and start earning money today!

anish2305/
airnote-hinglish-stt-ggml

Name: airnote-hinglish-stt-ggml
Availability: InStock

Oriserve/Whisper-Hindi2Hinglish-Swift

finetune

apache-2.0

Views

No views yet

AirNote Hinglish STT — GGML (whisper.cpp)

GGML conversions of Oriserve/Whisper-Hindi2Hinglish-Swift (a Whisper-base fine-tune that transcribes Hindi/Hinglish speech to romanized Hinglish) for on-device inference with whisper.cpp — no Python, no PyTorch.

Used by AirNote for fully native, offline dictation on macOS and Windows.

Files

File	Precision	Size	Notes
`ggml-oriserve-hinglish-fp16.bin`	fp16	148 MB	Recommended. Faithful to the original (matches the PyTorch model).
`ggml-oriserve-hinglish-q8_0.bin`	q8_0	78 MB	Near-lossless, smaller.
`ggml-oriserve-hinglish-q5_0.bin`	q5_0	53 MB	Smallest; minor accuracy drift on short clips.

Usage

whisper-cli -m ggml-oriserve-hinglish-fp16.bin -f audio.wav -l hi

Input: 16 kHz mono PCM. Output: romanized Hinglish.

Conversion recipe

Converted from the original model.safetensors with whisper.cpp's models/convert-h5-to-ggml.py (fp16 default), then quantized with whisper-quantize (q8_0 / q5_0). The checkpoint is a standard WhisperForConditionalGeneration (Whisper-base: d_model 512, 6+6 layers, 80 mel bins), so it converts cleanly.

Credit & license

All model weights and training are by Oriserve — see the original model card. This repo only redistributes format conversions (GGML) under the same Apache-2.0 license.

anish2305/airnote-hinglish-stt-ggml