gtm-hi2hing-ggml

GTypeMotion Hindi2Hinglish (GGML / whisper.cpp)

GGML-converted version of Oriserve/Whisper-Hindi2Hinglish-Apex for use with whisper.cpp in the GTypeMotion desktop app.

What this is

This is a Q8_0 quantized GGML build of Oriserve's Hindi→Hinglish speech recognition model, packaged for direct use with whisper.cpp. It produces Romanized Hinglish output (e.g. "Kya ab reuse bhi EIC banegi?"), not Devanagari.

Files

• ggml-apex-hinglish-q8_0.bin — Q8_0 quantized, 834 MB, recommended
• For smaller download, see Marquestra's Apex-GGML repo which also has F16 (1.5 GB) and Q5_0 (547 MB) builds.

Use with whisper.cpp

whisper-cli \
    -m ggml-apex-hinglish-q8_0.bin \
    -f audio.wav \
    -l auto \
    --print-progress

For best results on short / noisy clips, use the Silero VAD:

whisper-cli \
    -m ggml-apex-hinglish-q8_0.bin \
    -f audio.wav \
    -l auto \
    --vad --vad-model ggml-silero-v5.1.2.bin

Audio must be 16 kHz mono. Convert with ffmpeg if needed:

ffmpeg -i input.mp4 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav

Model details

• Base model: Oriserve/Whisper-Hindi2Hinglish-Apex
• Architecture: OpenAI Whisper large-v3-turbo (32-layer encoder, 4-layer decoder, 128 mel bins)
• Training: ~550 hours of noisy Indian-accented Hindi speech, fine-tuned for Hinglish output
• Quantization: Q8_0 (lossless vs F16 on tested clips)
• License: Apache 2.0 (inherited from base model)

Use with GTypeMotion

The GTypeMotion desktop app downloads this file automatically on first launch. No setup needed — just open the app and click "Download model".

Credits

• Oriserve for the original fine-tune
• Marquestra for the initial GGML conversion
• whisper.cpp for the inference engine