Speech to Text in Python

Local audio transcription with OpenAI Whisper. No API keys, no cloud, no data leaving your machine.

Why

You have an audio file you want transcribed. The usual options:

• OpenAI / AssemblyAI / Deepgram APIs - fast, accurate, but your audio goes to a server.
• macOS dictation / Otter.ai - also cloud.
• This - Whisper running entirely on your machine. Sensitive audio (interviews, medical notes, legal recordings) never leaves.

Requirements

• Python 3.7+
• FFmpeg (for audio decoding)
• ~1-10 GB disk space depending on which Whisper model you load

Install

pip install openai-whisper ffmpeg-python torchaudio

Install FFmpeg if you don't have it:

# macOS
brew install ffmpeg

# Ubuntu / Debian
sudo apt install ffmpeg

# Windows (with Chocolatey)
choco install ffmpeg

Verify:

ffmpeg -version

Usage

Edit text-to-speech.py to point at your audio file:

audio_path = "/path/to/your/audio.wav"

Then run:

python text-to-speech.py

Output: transcript printed to terminal and saved as a .txt file alongside the audio.

Models

Model	Speed	Accuracy	RAM
tiny	Fastest	Lowest	~1 GB
base	Fast	Good	~1 GB
small	Medium	Better	~2 GB
medium	Slow	Better	~5 GB
large	Slowest	Best	~10 GB

Change the model in the script:

model = whisper.load_model("small")  # or "base", "medium", "large"

The first time you load a model, Whisper downloads it from the OpenAI CDN.

Supported Audio Formats

Anything FFmpeg can decode: WAV, MP3, M4A, FLAC, OGG, AAC, WMA, etc.

Troubleshooting

FFmpeg errors - verify the install with ffmpeg -version. On Windows, make sure C:\ffmpeg\bin is in your PATH.

FileNotFoundError - use an absolute path to the audio file. Check for typos in the filename.

Model download fails - first run requires internet to fetch the model. Check disk space too (models range from ~150 MB to ~3 GB).

Poor accuracy - try a larger model. Accuracy on noisy or accented audio scales heavily with model size. The large model handles non-English audio much better than base.

License

MIT