Table of Contents
Whisper Large-v3-Turbo (released late 2024) trades a small quality drop for a ~4× speed gain. For batch transcription that's a huge win; for highest-quality multilingual it's the wrong choice.
TL;DR
Whisper Large-v3-Turbo is ~4× faster than Large-v3 with ~1-2% WER increase on standard benchmarks. Use Turbo for batch / English-heavy / speed-critical workloads. Use full Large-v3 for highest-quality multilingual or low-resource languages.
Comparison
| Aspect | Large-v3 | Large-v3-Turbo |
|---|---|---|
| Parameters | 1.55B | 809M |
| VRAM (FP16) | ~6 GB | ~3.5 GB |
| Speed (RTX 5090) | ~9× RTF | ~36× RTF |
| English WER | ~7.5% | ~7.7% |
| Multilingual WER (avg) | ~12.5% | ~14.0% |
| Low-resource lang quality | Strong | Drops noticeably |
Verdict
- Batch transcription, English: Turbo
- Multi-language production: Large-v3
- Real-time captioning, English: Turbo (lower latency)
- Voice agent: either works; Turbo if VRAM-tight
Bottom line
Turbo for speed, Large-v3 for quality. See Whisper hosting.
