Transcription
Whisper Large V3 Turbo
Whisper Large V3 TurboThe most performant Whisper Large V3 Turbo implementation, achieving 2400x real-time factor (1 hour of audio transcribed in 1.5 seconds).
Model details
Transcribe audio files at up to 2400x real-time factor (RTF). To add streaming or diarization get in touch with our engineers. Learn more about our optimized Whisper transcription pipeline in our launch blog.
✕
Accuracy vs latency — semantic word-error-rate vs median time-to-final-segment (single stream), plotted against closed-source providers.Recommended setups for different use cases:
Balanced
GPU type: H100 MIG
Concurrency target: <= 18
Highly latency-sensitive
GPU type: H100
Concurrency target: <= 32
Cost-sensitive
GPU type: L4
Concurrency target: <= 12
Try the example code below, or check out our API documentation for more detailed information.
Input
1import requests
2import os
3
4model_id = "" # Add the model ID for your deployment
5
6# Read secrets from environment variables
7baseten_api_key = os.environ["BASETEN_API_KEY"]
8
9# Define the request payload
10payload = {
11 "whisper_input": {
12 "audio": {
13 "url": "https://example.com/audio.wav"
14 # "audio_b64": "BASE64_ENCODED_AUDIO" # Uncomment if using Base64
15 },
16 "whisper_params": {
17 "prompt": "Optional transcription prompt",
18 "audio_language": "en"
19 }
20 }
21}
22
23resp = requests.post(
24 f"https://model-{model_id}.api.baseten.co/environments/production/predict",
25 headers={"Authorization": f"Api-Key {baseten_api_key}"},
26 json=payload
27)
28
29print(resp.json())JSON output
1{
2 "language_code": "en",
3 "language_prob": null,
4 "segments": [
5 {
6 "text": "That's one small step for man, one giant leap for mankind.",
7 "log_prob": -0.8644908666610718,
8 "start_time": 0,
9 "end_time": 9.92
10 }
11 ]
12}
