Model ID
cai-whisper-large-v3-turbo-slim
Whisper Large V3 Turbo Slim by CompactifAI
Model Overview
Section titled “Model Overview”Base Architecture
Whisper Large V3 Turbo
Model Highlights
Section titled “Model Highlights”- Quantum-inspired compression delivers ~50% model size reduction while maintaining strong accuracy. Average WER change is -0.40 pp overall, with +0.68 pp in English and -1.83 pp in Spanish versus the full turbo baseline, offering strong accuracy per dollar for production speech workloads.
- Significantly faster inference: compared to the full turbo baseline, the slim model is ~2.5× faster in Real Time Factor (RTF), enabling quicker transcription and higher throughput.
- Excellent performance in English and Spanish, delivering high-quality ASR results.
- Optimized for long-form audio and robust to background noise and accents.
Specifications
Section titled “Specifications”| Specification | Value |
|---|---|
| Parameters | ~0.4B |
| Log-Mel | spectrogram with 128 Mel bins, computed at a 16 kHz sampling rate, using an FFT window of 400 samples (~25 ms) and a hop length of 160 samples (~10 ms) |
| Max Audio Duration | ~30 seconds per chunk (sliding window supported) |
| Supported file types | flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, webm |
Recommended Use Cases
Section titled “Recommended Use Cases”- High-fidelity meeting and podcast transcription.
- Real-time captioning with batch segmentation.
- Multilingual customer-support call analytics.