OpenAI Whisper API: Complete Integration Guide

What Whisper Does

Whisper transcribes audio to text with state-of-the-art accuracy. Supports 57 languages and can translate non-English audio directly to English text.

Cost: $0.006 per minute - extremely affordable for most use cases.

Basic Setup

python
from openai import OpenAI
client = OpenAI(api_key='your-key')
Transcribe an audio file
with open('meeting.mp3', 'rb') as audio_file:
    transcript = client.audio.transcriptions.create(
        model='whisper-1',
        file=audio_file
    )
print(transcript.text)

Language Detection and Translation

python
Auto-detect language
transcript = client.audio.transcriptions.create(
    model='whisper-1',
    file=audio_file,
    response_format='verbose_json'  # Includes detected language
)
print(f'Language: {transcript.language}')
print(f'Text: {transcript.text}')
Translate non-English to English
translation = client.audio.translations.create(
    model='whisper-1',
    file=spanish_audio
)
print(translation.text)  # Always English output

Timestamps and Segmentation

python
transcript = client.audio.transcriptions.create(
    model='whisper-1',
    file=audio_file,
    response_format='verbose_json',
    timestamp_granularities=['segment', 'word']  # Both segment and word timestamps
)
for segment in transcript.segments:
    print(f'[{segment.start:.1f}s - {segment.end:.1f}s]: {segment.text}')

Cost Optimization

python
import librosa
import soundfile as sf
def optimize_audio_for_whisper(input_path: str, output_path: str):
    # Load and resample to 16kHz mono (Whisper native format)
    audio, sr = librosa.load(input_path, sr=16000, mono=True)
    
    # Trim silence (saves significant cost on meetings with long pauses)
    audio_trimmed, _ = librosa.effects.trim(audio, top_db=20)
    
    # Save as 16-bit PCM WAV (smaller than MP3 for short clips)
    sf.write(output_path, audio_trimmed, 16000, subtype='PCM_16')
    
    original_duration = librosa.get_duration(filename=input_path)
    trimmed_duration = len(audio_trimmed) / 16000
    savings = (original_duration - trimmed_duration) / original_duration
    print(f'Audio reduced by {savings:.1%}')
    return output_path

Handling Large Files

Whisper API has a 25MB file size limit. For longer audio:

python
from pydub import AudioSegment
def transcribe_long_audio(file_path: str, chunk_minutes: int = 10) -> str:
    audio = AudioSegment.from_file(file_path)
    chunk_ms = chunk_minutes * 60 * 1000
    chunks = [audio[i:i+chunk_ms] for i in range(0, len(audio), chunk_ms)]
    
    transcripts = []
    for i, chunk in enumerate(chunks):
        chunk_path = f'/tmp/chunk_{i}.mp3'
        chunk.export(chunk_path, format='mp3', bitrate='64k')
        
        with open(chunk_path, 'rb') as f:
            result = client.audio.transcriptions.create(
                model='whisper-1', file=f
            )
        transcripts.append(result.text)
    
    return ' '.join(transcripts)

Real-World Applications

Meeting transcription: Record meetings, transcribe, then use GPT-4o to extract action items and summaries.

Customer service analytics: Transcribe support calls to identify common issues and sentiment patterns.

Subtitle generation: Whisper with word timestamps generates accurate SRT subtitle files.

Multilingual support: Support users in any of 57 languages without separate language-specific models.

Quality Tips

16kHz mono audio gives best results (Whisper native format)

Use verbose_json format to detect and handle low-confidence segments

For specialized vocabulary (medical, legal, technical), use the prompt parameter to provide context

Background noise significantly impacts accuracy - consider noise reduction preprocessing