CoolFace
Modelpublic

Marquestra/Whisper-Hindi2Hinglish-Apex-GGML

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
0likes
Model Card

Whisper-Hindi2Hinglish-Apex — GGML (whisper.cpp)

GGML builds of `Oriserve/Whisper-Hindi2Hinglish-Apex` for use with whisper.cpp.

Apex is a fine-tune of `openai/whisper-large-v3-turbo` (32-layer encoder / 4-layer decoder, 128 mel bins) that transcribes Hindi/English/Hinglish speech and outputs romanized Latin ("Kaise ho bhai", not Devanagari). These are on-device, CPU/Metal-friendly GGML weights for real-time dictation.

Files

FileQuantSizeNotes
ggml-apex-hinglish-fp16.binfp161.5 GBReference / highest fidelity
ggml-apex-hinglish-q8_0.binq8_0834 MBRecommended — lossless vs fp16 on our eval, fits <1 GB
ggml-apex-hinglish-q5_0.binq5_0547 MBSmallest; minor drift vs fp16

How these were produced

Converted with the official whisper.cpp converter, then quantized with the whisper.cpp quantize tool.

bash
# 1) convert HF (safetensors) -> fp16 GGML
git clone https://github.com/openai/whisper           # supplies whisper/assets/mel_filters.npz (mel_128)
git clone https://github.com/ggml-org/whisper.cpp
git clone https://huggingface.co/Oriserve/Whisper-Hindi2Hinglish-Apex
python3 ./whisper.cpp/models/convert-h5-to-ggml.py ./Whisper-Hindi2Hinglish-Apex ./whisper .
# -> ggml-model.bin  (renamed to ggml-apex-hinglish-fp16.bin)

# 2) quantize
./whisper.cpp/build/bin/quantize ggml-apex-hinglish-fp16.bin ggml-apex-hinglish-q8_0.bin q8_0
./whisper.cpp/build/bin/quantize ggml-apex-hinglish-fp16.bin ggml-apex-hinglish-q5_0.bin q5_0

Apex ships BF16 weights; because NumPy has no bfloat16 dtype, the one line data = list_vars[src].squeeze().numpy() in convert-h5-to-ggml.py is changed to ... .squeeze().float().numpy() (a lossless BF16→FP32 upcast; the script then writes fp16 as usual).

Verification (loads as "large v3" — turbo has a 4-layer decoder)

Header on load: n_vocab = 51866, n_mels = 128, type = large v3, ftype = fp16 — matching stock large-v3-turbo (no vocab/mel corruption).

On a 24-clip Hindi/English/Hinglish set:

  • —q8_0 output is identical to fp16 on every clip (near-lossless int8).
  • —Output stays romanized (0 Devanagari codepoints).

Usage (whisper.cpp)

bash
./build/bin/whisper-cli -m ggml-apex-hinglish-q8_0.bin -f audio.wav -l auto
# VAD-gate short/near-silent clips to avoid degenerate output:
./build/bin/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. Very short (<1 s) / near-silent clips can produce degenerate output on turbo models — gate them with Silero VAD (as above).

License

Apache-2.0, inherited from the base model Oriserve/Whisper-Hindi2Hinglish-Apex and openai/whisper-large-v3-turbo.

Credits