Whisper AI: Complete Transcription Guide

OpenAI Whisper Transcription Guide

Whisper is OpenAI's speech recognition model that delivers remarkable accuracy across 99 languages. Whether you're transcribing interviews, meetings, podcasts, or videos, this guide covers everything you need to know.

What is Whisper?

Whisper is an automatic speech recognition (ASR) model trained on 680,000 hours of multilingual audio. It offers:

• 99 language support
• Translation to English
• Punctuation and formatting
• Speaker-aware transcription
• Noise robustness

Ways to Use Whisper

1. OpenAI API (Easiest)

• No setup required
• Pay per minute
• Best for occasional use
• 25MB file limit

2. Local Installation (Free)

• Requires Python setup
• Free after setup
• Best for high volume
• No file size limits

3. Apps Using Whisper (Most Convenient)

• User-friendly interfaces
• Various pricing models
• Additional features

Method 1: OpenAI API

Setup:

python
pip install openai

Basic Usage:

python
from openai import OpenAI
client = OpenAI()
audio_file = open("meeting.mp3", "rb")
transcript = client.audio.transcriptions.create(
    model="whisper-1",
    file=audio_file
)
print(transcript.text)

With Options:

python
transcript = client.audio.transcriptions.create(
    model="whisper-1",
    file=audio_file,
    response_format="srt",  # or vtt, json, text
    language="en",  # ISO code
    prompt="Technical meeting about AI"  # Context helps accuracy
)

Pricing:

• $0.006 per minute
• Example: 1-hour video = $0.36

Method 2: Local Installation

Requirements:

• Python 3.8+
• FFmpeg
• 4-10GB RAM (depending on model)

Install:

bash
pip install openai-whisper

Install FFmpeg:

• Mac: brew install ffmpeg
• Windows: Download from ffmpeg.org
• Linux: apt install ffmpeg

Basic Usage:

bash
whisper audio.mp3 --model medium

Model Sizes:

| Model | Size | RAM | Speed | Accuracy | |-------|------|-----|-------|----------| | tiny | 39M | ~1GB | Fastest | Basic | | base | 74M | ~1GB | Fast | Good | | small | 244M | ~2GB | Medium | Better | | medium | 769M | ~5GB | Slow | Great | | large | 1550M | ~10GB | Slowest | Best |

Recommended: Start with medium for good balance.

Advanced Options:

bash
whisper interview.mp3 \
  --model medium \
  --language English \
  --output_format srt \
  --output_dir ./transcripts \
  --task transcribe  # or translate

Method 3: Apps Using Whisper

MacWhisper (Mac)

• Beautiful Mac app
• Drag-and-drop
• Free tier available
• Pro: $29 one-time

Whisper Transcription (iOS)

• Mobile transcription
• On-device processing
• Privacy-focused

TurboScribe

• Web-based
• Unlimited in pro tier
• Speaker labels
• From $10/month

Descript

• Full audio/video editor
• Uses Whisper + custom models
• From $15/month

Otter.ai

• Real-time transcription
• Meeting integration
• Free tier available

Tips for Best Results

1. Audio Quality

• Use good microphone
• Minimize background noise
• Record in quiet environment
• Higher sample rate = better

2. Use Prompts Provide context to improve accuracy:

python
prompt="Meeting about machine learning, discussing neural networks, transformers, and GPT models. Speakers: John (CEO), Sarah (CTO)."

3. Choose Right Model

• Quick notes → tiny/base
• Important content → medium
• Critical accuracy → large

4. Post-Processing Even Whisper makes mistakes. Review:

• Proper nouns and names
• Technical terminology
• Numbers and dates
• Homophone errors

Common Use Cases

Meeting Notes:

• Record meeting (Zoom, Teams, etc.)
• Export audio
• Run through Whisper
• Feed to ChatGPT: "Summarize these meeting notes and extract action items"

Podcast Transcription:

• Download episode
• Transcribe with Whisper (large model)
• Edit transcript
• Publish as show notes

Video Subtitles:

bash
whisper video.mp4 --model medium --output_format srt

This creates video.srt for subtitles.

Lecture Notes:

• Record lecture
• Transcribe
• Use ChatGPT to create summary and study notes

Handling Long Files

For files over 25MB (API limit):

Option 1: Split Audio

bash
ffmpeg -i long_file.mp3 -f segment -segment_time 600 -c copy output%03d.mp3

This splits into 10-minute chunks.

Option 2: Local Whisper No file size limits with local installation.

Option 3: Use Apps Most apps handle long files automatically.

Language Translation

Whisper can translate any supported language to English:

bash
whisper spanish_audio.mp3 --task translate

Or via API:

python
translation = client.audio.translations.create(
    model="whisper-1",
    file=audio_file
)

Accuracy Comparison

In testing, Whisper's accuracy on clean audio:

• English: 95-98%
• Major languages: 90-95%
• With noise: 80-90%
• Heavy accents: 85-95%

Factors affecting accuracy:

• Audio quality
• Background noise
• Multiple speakers
• Speaking speed
• Technical vocabulary

Workflow Example

Journalist Interview:

The combination of Whisper + AI editing can cut transcription time by 80% compared to manual transcription.