OpenAI released Whisper in 2022 with five model sizes. In 2024, they added
the large-v3 model with significant accuracy improvements. In 2026,
the large-v3-turbo model joined the lineup with the accuracy of v3
at twice the speed.
Choosing the right Whisper model matters more in 2026 than in 2024 because the speed-quality tradeoff has shifted. The default "use the largest model" advice no longer holds. Here's the breakdown.
| Model | Size | VRAM | Speed (1h audio) | Accuracy (WER) |
|---|---|---|---|---|
| tiny | 39M params | 1 GB | 2 min | 7.6% |
| base | 74M params | 1 GB | 4 min | 5.0% |
| small | 244M params | 2 GB | 7 min | 3.4% |
| medium | 769M params | 5 GB | 15 min | 2.5% |
| large-v2 | 1.5B params | 10 GB | 30 min | 2.1% |
| large-v3 | 1.5B params | 10 GB | 30 min | 1.8% |
| large-v3-turbo | 0.8B params | 6 GB | 15 min | 1.9% |
Speed measured on an RTX 4090 with faster-whisper backend. WER
(Word Error Rate) measured on the LibriSpeech test-clean English set. Lower WER
is better.
large-v3-turbo is the headline of 2026. It distills the v3 model into a smaller architecture (0.8B vs 1.5B parameters) without losing meaningful accuracy. The result: near-v3 accuracy at 2x the speed and 60% of the VRAM.
The turbo model uses a simplified decoder (4 transformer layers instead of 32), which is why it's faster. For pure transcription, this trade-off is a clear win. For tasks that need the decoder's full reasoning (translation, long-form context), large-v3 still wins.
Use large-v3-turbo. It's the best speed/quality tradeoff for non-real-time use. 15 minutes for an hour of audio is fast enough that you won't notice the wait.
Use small or medium. Real-time requires faster-than-realtime inference. Large models can't keep up on CPU; medium processes 1 hour of audio in 15 minutes, fast enough for a 2x latency buffer.
Use large-v3 if accuracy is critical (legal, medical). Use large-v3-turbo if you need throughput. The 0.1% accuracy difference becomes meaningful at scale.
Use base or small. Larger models on CPU take hours per file. base is the sweet spot for CPU-only laptops.
Use tiny. Whisper.cpp can run tiny on modern phones in real-time. The accuracy is poor but usable for voice memos.
Use large-v3 or large-v3-turbo. The translation task needs the full decoder reasoning. Turbo works fine for short utterances; large-v3 is better for long-form translation.
The Word Error Rate differences look small (1.8% vs 5.0%), but they translate to very different user experiences:
| Model | 1 hour of audio | Real-world experience |
|---|---|---|
| tiny | ~140 errors | Misunderstands technical terms, accents. Needs heavy editing. |
| base | ~90 errors | Works for casual content. Fails on names, jargon. |
| small | ~60 errors | Good for podcasts, interviews. Needs light editing. |
| medium | ~45 errors | Production-ready for most content. Few errors. |
| large-v3 | ~32 errors | Ship-ready. Suitable for broadcast, legal, medical. |
| large-v3-turbo | ~34 errors | Ship-ready for 95% of use cases. Twice the speed. |
The model file is only half the story. The backend you run it on affects speed by 2-10x:
For self-hosting, faster-whisper is the standard in 2026. For API work, you're using OpenAI or Groq's hosted version.
You can quantize Whisper models to int8 or int4 to reduce VRAM and increase speed, at a small accuracy cost:
| Precision | VRAM reduction | Speed up | Accuracy loss |
|---|---|---|---|
| fp32 (default) | 1x | 1x | 0% |
| fp16 | 2x | 1.5x | < 0.1% |
| int8 | 4x | 2x | ~0.3% |
| int4 | 8x | 3x | ~1% |
fp16 is the sweet spot for GPUs. int8 is the sweet spot for CPU-only inference. int4 is rarely worth the accuracy loss.
| Your situation | Use this |
|---|---|
| Default for most users | large-v3-turbo |
| Real-time / streaming | medium |
| CPU only (no GPU) | base or small |
| Phone or embedded | tiny |
| Maximum accuracy, no GPU limit | large-v3 |
| Translation to other languages | large-v3 |
| Low VRAM (8GB GPU) | large-v3-turbo int8 |
pip install faster-whisper
from faster_whisper import WhisperModel
model = WhisperModel("large-v3-turbo", device="cuda", compute_type="float16")
segments, info = model.transcribe("audio.mp3", beam_size=5)
for segment in segments:
print(f"[{segment.start:.1f}s - {segment.end:.1f}s] {segment.text}")
Expected time on RTX 4070: 8-10 minutes for 1 hour of audio.
Use large-v3 with translation. Process 10 files in parallel on
an A100 80GB. Each file takes 4-6 minutes for transcription + translation.
Use medium with streaming chunks. 2-3 second latency on M2 Mac.
Output as WebVTT for browser display.
Self-host large-v3 with on-prem GPU. Audio never leaves your network. Use int8 quantization if GPU VRAM is limited.
No announcement yet. v3-turbo is the current frontier. If you need better accuracy today, use v3 or v3-turbo with post-processing.
Distil-Whisper is now superseded by v3-turbo for most use cases. Turbo is faster and more accurate.
Yes. OpenAI recommends it as the default in 2026 for any non-streaming use case. It's the same accuracy with 2x the speed.
For v3-turbo at fp16: 6GB minimum, 8GB recommended. For v3 at fp16: 10GB minimum. For larger models or lower precision, you can get away with less.
Yes, on the same GPU. A 24GB card can run v3-turbo + v3 simultaneously. Or process multiple audio files in parallel batches.
In 2026, the right default is large-v3-turbo. It gives you near-best accuracy at half the cost and twice the speed of v3. Only switch to v3 if you need the absolute best accuracy (legal, medical, broadcast) or translation to languages where turbo struggles.
If you're not sure, start with turbo. You can always re-transcribe with a larger model later if the output isn't good enough.