Filename	Quant	Size	Precision
`model_f16.gguf`	F16	631 MB	~fp16 (near-lossless)
`model_q8_0.gguf`	Q8_0	373 MB	8-bit quantized linear layers
`model_q4_0.gguf`	Q4_0	222 MB	4-bit quantized linear layers

Parity vs HuggingFace Reference

Tested on 10 audio samples from FLEURS hi_in.

Quant	Mean CER (vs HF)	Mean WER (vs HF)	Status
F16	0.06%	0.27%	PASS
Q8_0	0.63%	2.46%	PASS
Q4_0	6.37%	28.44%	FAIL

CER/WER measured against HF model output (not ground truth) to isolate quantization error.

Usage

C++ CLI

bash

# Clone and build
git clone --recursive https://github.com/liodon-ai/wav2vec2.cpp
cd wav2vec2.cpp && mkdir build && cd build && cmake .. && make -j

# Download and run
huggingface-cli download liodon-ai/Wav2Vec2-large-xlsr-hindi-GGUF model_q8_0.gguf --local-dir .
./wav2vec2-cli -m model_q8_0.gguf -f audio.wav

# With word timestamps
./wav2vec2-cli -m model_q8_0.gguf -f audio.wav -w

# SRT subtitle output
./wav2vec2-cli -m model_q8_0.gguf -f audio.wav --format srt > output.srt

Python (via subprocess)

python

import subprocess, json
result = subprocess.run(
    ['./wav2vec2-cli', '-m', 'model_q8_0.gguf', '-f', 'audio.wav', '--format', 'json'],
    capture_output=True, text=True
)
data = json.loads(result.stdout)
print(data['transcript'])

Source Model

Original: theainerd/Wav2Vec2-large-xlsr-hindi
Architecture: XLS-R
Language: hi
Task: Automatic Speech Recognition

Conversion

Converted using wav2vec2.cpp scripts/convert_to_gguf.py.

bash

python scripts/convert_to_gguf.py theainerd/Wav2Vec2-large-xlsr-hindi model_f16.gguf --dtype f16
python scripts/convert_to_gguf.py theainerd/Wav2Vec2-large-xlsr-hindi model_q8_0.gguf --dtype q8_0
python scripts/convert_to_gguf.py theainerd/Wav2Vec2-large-xlsr-hindi model_q4_0.gguf --dtype q4_0

License

Follows the license of the source model.