whisper-small-quantized.w4a16

Model Overview

• Model Architecture: whisper-small
  • Input: Audio-Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT4
• Release Date: 04/16/2025
• Version: 1.0
• Model Developers: Neural Magic

Quantized version of openai/whisper-small.

Model Optimizations

This model was obtained by quantizing the weights of openai/whisper-small to INT4 data type, ready for inference with vLLM >= 0.5.2.

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from vllm.assets.audio import AudioAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
 model="neuralmagic/whisper-small-quantized.w4a16",
 max_model_len=448,
 max_num_seqs=400,
 limit_mm_per_prompt={"audio": 1},
)

# prepare inputs
inputs = { # Test explicit encoder/decoder prompt
 "encoder_prompt": {
 "prompt": "",
 "multi_modal_data": {
 "audio": AudioAsset("winning_call").audio_and_sample_rate,
 },
 },
 "decoder_prompt": "<|startoftranscript|>",
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.0, max_tokens=64))
print(f"PROMPT : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")

vLLM also supports OpenAI-compatible serving. See the documentation for more details.

Creation

This model was created with llm-compressor by running the code snippet below.

Evaluation

The model was evaluated on LibriSpeech and Fleurs datasets using lmms-eval, via the following commands:

Benchmark	Split	BF16	W4A16	Recovery (%)
LibriSpeech (WER)	test-clean	3.4435	3.5246	97.70%
	test-other	7.4884	7.9422	94.30%
Fleurs (X→en, WER)	cmn_hans_cn	23.0642	25.5212	90.37%
	en	6.2002	6.5092	95.25%
	yue_hant_hk	16.2557	22.2870	73.93%

Base model

openai/whisper-small

(219)

Collection of quantized whisper models created by OpenAI • 19 items • Updated Apr 30 • 4