Whisper Turbo v3 Self-Hosted

Whisper Turbo (large-v3-turbo) is a distilled variant of Whisper large-v3 with the decoder layer count cut from 32 to 4. Roughly 8x faster transcription at nearly identical accuracy on most languages. On our dedicated GPU hosting it is the default self-hosted transcription model in 2026.

Contents

• VRAM
• Deployment
• Speed comparison
• Quality caveats

VRAM

~1.6 GB at FP16. Runs on any card including the 3050. Whisper is small compared to an LLM – any dedicated card has the capacity.

Deployment

faster-whisper is the recommended runtime – CTranslate2 backend, INT8 quantisation, batched inference:

pip install faster-whisper

from faster_whisper import WhisperModel
model = WhisperModel("turbo", device="cuda", compute_type="float16")
segments, info = model.transcribe("audio.wav", beam_size=5)
for s in segments:
    print(s.text)

For an HTTP service, wrap in FastAPI or use a pre-built server like whisper-webservice.

Speed

Transcribing one hour of audio:

Quality

Turbo matches large-v3 on English and major European languages. On low-resource languages (Swahili, Burmese, Telugu) accuracy drops slightly. For production English workloads, always pick Turbo over large-v3. For rare languages, test both.

See Whisper + diarization for speaker separation.