Audio Transcription in 3 Minutes per Hour of Recording, Powered by Whisper on GPU

The Whisper model has a known weakness: when a recording contains silence, music or noise, it can fill that spot with a sentence that sounds natural and was never spoken. In studies on hundreds of thousands of recordings with no speech, this happened in four cases out of ten. For meeting notes that is annoying. For a record of a customer call or a medical consultation it is a problem that is hard to undo later.

So every recording first goes through speech detection, and only the parts where someone is actually talking reach the model. Repeating loops are removed, and sentences from the list of typical fabrications are cut automatically. The report for every transcript shows how many seconds of silence and noise we filtered out, so you know the text matches what was in the recording.

On public test sets Whisper reads Polish with an error rate of a few percent. On real recordings from a doctor's office the same model, in a study from AGH University in Kraków, got every second word wrong. The gap comes from the fact that benchmarks are clear, read sentences, while real conversations have abbreviations, proper names, numbers and two people talking at once.

That is why before rollout we ask for an hour of typical recordings from your company. We transcribe them, compare the result with a manual transcript and hand you a report with accuracy for each type of recording: phone calls, meetings, dictation. On that basis we pick the engine and the settings, and you know what to expect before you pay the first invoice.

A general model does not know drug names, the item numbers from your catalogue or the abbreviations your team uses. The usual fix is fine-tuning on recordings, which means collecting and transcribing many hours of audio. There is a simpler way. The model can learn vocabulary from text alone: from documentation, glossaries, meeting minutes and earlier transcripts you already have.

We run the training on our own GPUs in Poland, and the tuned model works for you alone. We agree the scope and cost after looking at your materials, because both depend on how much specialist vocabulary the model needs to see.