Whisper — OpenAI Speech-to-Text

What it is

Whisper is OpenAI's open-source speech recognition model. It transcribes audio and video files to text with high accuracy across 99 languages. The model runs locally, requires no API key, and produces output in plain text, SRT subtitles, VTT, JSON, or TSV formats.

Whisper targets developers, content creators, and researchers who need reliable transcription without sending audio to cloud services. Multiple model sizes (tiny to large) trade accuracy for speed.

How it saves time or tokens

Manual transcription takes 4-6x the audio duration. Whisper transcribes a 1-hour podcast in minutes on a modern GPU or 15-30 minutes on CPU. The output includes word-level timestamps, making it directly usable for subtitle generation.

For AI workflows, Whisper converts audio content into text that LLMs can process. Meeting recordings, podcast episodes, and lecture videos become searchable, summarizable text.

How to use

1. Install Whisper:

pip install openai-whisper

1. Transcribe audio from the command line:

whisper audio.mp3 --model medium --language en --output_format srt

1. Use the Python API for programmatic access:

import whisper

model = whisper.load_model('medium')
result = model.transcribe('audio.mp3')
print(result['text'])

Example

import whisper

model = whisper.load_model('medium')

# Transcribe with word-level timestamps
result = model.transcribe(
    'meeting.mp3',
    word_timestamps=True,
    language='en'
)

# Access segments with timestamps
for segment in result['segments']:
    print(f"[{segment['start']:.1f}s - {segment['end']:.1f}s] {segment['text']}")

# Output:
# [0.0s - 3.2s] Welcome to the quarterly review.
# [3.2s - 7.8s] Let us start with the revenue numbers.

Related on TokRepo

• AI Tools for Voice -- Speech synthesis and recognition tools
• AI Tools for Content -- Content creation and processing tools

Common pitfalls

• The large model requires 10GB+ of VRAM. Use the medium or small model if GPU memory is limited. CPU inference works but is significantly slower.
• Whisper hallucinates on silent audio segments, sometimes generating repetitive or nonsensical text. Pre-process audio to trim silence.
• Non-English transcription accuracy varies by language. Languages with less training data produce lower quality output.