Choose local transcription for your own recording
Choose local transcription for your own recording — Free local transcription with Whisper and FFmpeg. What actually stays local; Model size, memory and execution cost; Define a finished-transcript criterion. Practical assignment: Choose local transcription for your own recording.
Demo edition. The materials are undergoing editorial review.
Topics covered
- What actually stays local
- Model size, memory and execution cost
- Define a finished-transcript criterion
What actually stays local
Local transcription means the model processes audio on a computer you control. This course uses whisper.cpp, a C/C++ inference implementation for Whisper models. Initial software and model downloads need networking, but prepared audio can then be transcribed without it. This differs from uploading a recording to a website containing Whisper in its name. Record your own one-minute note with a date, filename and number; do not use someone else’s private conversation. Create originals, work and reviewed directories. Preserve the original and keep working WAV files and raw text separate. Running locally does not encrypt your disk or prevent a folder from syncing to the cloud. Check storage location, permissions and backup settings. Privacy depends on the actual data route rather than the program’s name. Obtain appropriate permission before including other people in a recording.
Model size, memory and execution cost
Start with multilingual base for Ukrainian, Russian or Italian audio. The .en suffix identifies an English-only model; base.en is not a smaller universal substitute. A larger model may help with difficult audio but requires more memory and time, so evaluate it on your own short sample. Memory tables for original PyTorch Whisper and whisper.cpp describe different implementations and are not interchangeable. Record RAM, available disk, CPU/GPU type, software version and model. Identical model files do not guarantee identical speed on different computers. MIT licensing for code and weights does not provide free GPUs, VPS hosting or APIs. This local route avoids remote per-call API billing but still consumes hardware, electricity and review time. If memory is insufficient, first reduce model size and concurrency while keeping the test sample unchanged for comparison.
Define a finished-transcript criterion
Raw ASR text is a draft of what the model recognized. It can omit a negation, change a number or add speech during silence. Before running your one-minute note, record the date, number and filename you actually spoke as reference facts. A finished transcript reproduces them after listening, marks unintelligible passages and links to the original. Do not silently turn transcription into paraphrase: removal of repetitions or grammatical cleanup belongs in a separate edited version. Automatic timestamps need checking too; they do not establish speaker identity. Standard Whisper is not reliable identification of who a person is. The course deliverable is reviewed transcript.txt, raw output and a settings log. If the software was not run, submit a preparation plan rather than claiming completed transcription. Preserve uncertainty instead of guessing an inaudible word.
Practical assignment
Choose local transcription for your own recording
Continue to learning