OpenAI Whisper Large v3 Integration for Speech Recognition Whisper Large v3 is OpenAI's current flagship model for ASR, released in November 2023. Compared to Large v2, it reduced WER by 10-20% in most languages. On clear Russian audio, it reduced WER by 6-9%, and on telephony, it reduced WER by 15-20%. ### Key improvements v3 vs v2 - Trained on a wider range of languages with improved data - Fewer hallucinations in silence and noise - Better punctuation out of the box - Improved code-switching processing ### Infrastructure requirements For comfortable real-time performance, a GPU with ≥10 GB VRAM is required. The optimal choice is NVIDIA A10G or RTX 4090. The model runs on CPU, but at 0.1-0.3x real-time speed — only for offline tasks.
Using faster-whisper with int8 quantization, the model fits into 6–7 GB of VRAM at 1.5–2x real-time speed:```bash
pip install faster-whisper
```python
from faster_whisper import WhisperModel
model = WhisperModel(
"large-v3",
device="cuda",
compute_type="int8_float16"
)
segments, info = model.transcribe(
"meeting.wav",
language="ru",
vad_filter=True,
vad_parameters={"min_silence_duration_ms": 500}
)
```### Use cases: - Meeting and interview transcription - Automatic video captioning - Archival processing of call center audio databases. Integration via the OpenAI API (without self-hosting) takes 1 day. Self-hosted with optimization for specific hardware: 3–5 days.