← Back to Articles

OpenAI Whisper Model Benchmarks: CPU vs GPU Latency & Memory Footprint

By Jay Parmar & Raj Prajapati Published Jul 30, 2026 7 min read

1. Benchmark Methodology

Selecting the optimal Whisper model size for real-time speech translation involves balancing transcription accuracy against inference latency and system RAM/VRAM consumption. To provide empirical guidance for Voxxwire users, we conducted rigorous benchmarks across four Whisper model variants (Tiny, Base, Small, Medium) executing under faster-whisper with CTranslate2 INT8 quantization.

Test Environment Hardware:

2. Full Benchmark Comparison Table

Model Variant Parameters RAM / VRAM (INT8) CPU Latency (5s chunk) GPU Latency (5s chunk) Avg WER
Whisper Tiny 39 M ~ 150 MB 120 ms 35 ms 10.4%
Whisper Base 74 M ~ 290 MB 210 ms 55 ms 7.8%
Whisper Small 244 M ~ 780 MB 480 ms 110 ms 5.2%
Whisper Medium 769 M ~ 2.1 GB 1,420 ms 260 ms 4.1%
Advertisement

3. CPU (AVX-512) vs GPU (CUDA INT8) Performance

Our benchmarks demonstrate that INT8 quantization reduces memory consumption by approximately 60% compared to standard FP32 weights, enabling even Whisper Small to operate comfortably inside 1 GB of RAM.

On modern CPUs equipped with AVX-2 or AVX-512 instructions, Whisper Base achieves a Real-Time Factor (RTF) of 0.04 (meaning 5 seconds of spoken audio is transcribed in just 210 milliseconds). On NVIDIA GPUs via CUDA, latency drops to an astonishing 55ms.

4. Word Error Rate (WER) vs Real-Time Factor (RTF)

While Whisper Medium provides superior recognition accuracy for complex medical or technical vocabulary, its CPU latency (1.4 seconds) can create slight visual delay during live subtitles. Therefore, Voxxwire recommends Whisper Base or Whisper Small as the ideal sweet spot for real-time meeting translation on typical laptop hardware.

5. Hardware Recommendation Matrix

📊

Written by Jay Parmar & Raj Prajapati

Core maintainers of Voxxwire — dedicated to building privacy-first, open-source offline AI desktop applications.