README.md

---
license: apache-2.0
datasets:
- Yehor/cv10-uk-testset-clean-punctuated
language:
- uk
base_model:
- openai/whisper-large-v3-turbo
---

# Quantized Whisper Large V3 Turbo with calibration on Ukrainian

Quantized it using https://pypi.org/project/llmcompressor/

Data used for calibration: https://huggingface.co/datasets/Yehor/cv10-uk-testset-clean-punctuated

How to quantize:
https://colab.research.google.com/drive/1TsCMxwq9kqsWV8jabihFN7J78RKgyvnD?usp=sharing

## Usage

Install required packages:

```
pip install vllm polars
```

Run inference:

```python
import io
import wave

import numpy as np
import polars as pl

from vllm import LLM, SamplingParams


def bytes_to_numpy(_bytes):
    with wave.open(io.BytesIO(_bytes), "rb") as wr:
        if (nc := wr.getnchannels()) != 1:
            raise ValueError(f"num_channels must be 1, got {nc}")
        if (sw := wr.getsampwidth()) != 2:
            raise ValueError(f"sample_width must be 2, got {sw}")

        audio_data = wr.readframes(wr.getnframes())

        return np.frombuffer(audio_data, dtype=np.int16).astype(np.float32) / 32768.0


llm = LLM(
    model="Yehor/whisper-large-v3-turbo-quantized-uk",
    max_model_len=448,
    max_num_seqs=400,
    gpu_memory_utilization=0.8,
    limit_mm_per_prompt={"audio": 1},
)

df = pl.read_parquet("hf://datasets/Yehor/cv10-uk-testset-clean/data/train-*.parquet")


for row in df.iter_rows(named=True):
    inputs = {
        "encoder_prompt": {
            "prompt": "",
            "multi_modal_data": {
                "audio": (bytes_to_numpy(row["audio"]["bytes"]), 16_000,),
            },
        },
        "decoder_prompt": "<|startoftranscript|><|uk|><|transcribe|><|notimestamps|>",
    }

    sampling_params = SamplingParams(
        temperature=1.0,
        top_p=1.0,
        max_tokens=200,
    )
    outputs = llm.generate(inputs, sampling_params)

    print(f"PROMPT  : {outputs[0].prompt}")
    print(f"TRANSCRIPTION: {row['transcription']}")
    print(f"PREDICTION: {outputs[0].outputs[0].text}")
    print("==========================================")
```