Getting the Best Transcription From Your Recording

Whisper is good at transcribing clear speech, but it is only as good as the recording you give it. A few habits around how you capture audio make a bigger difference than any model choice. These are the adjustments that actually move the needle.

Place the microphone close to the speaker

The single biggest factor is distance. A microphone near the speaker captures the voice strongly and the room weakly. A microphone across a conference table captures everything โ€” paper rustling, chair squeaks, distant traffic โ€” and the voice becomes just one layer in the mix. If people take turns speaking into one device, position it so each voice reaches it clearly.

Control background noise

Noise competes with speech. A fan, an air conditioner, traffic outside the window, or a keyboard being typed on during the meeting all appear in the spectrogram alongside the voice and make it harder for the model to tell which sound is the speech. Close the window, mute notifications, and ask people not to type or shuffle papers while someone is speaking.

Speak one at a time

When two people talk at once, their sound patterns overlap in the spectrogram and the model has to untangle them. It can, sometimes, but crosstalk is where most transcription errors come from. In a meeting, that means one person speaking at a time and finishing before the next begins โ€” a habit that improves the meeting itself, not just the transcript.

Mind the sample rate

This tool decodes audio at 16 kHz, which is the rate Whisper expects. Recordings made at unusual sample rates, or re-encoded several times through messaging apps, can arrive garbled or with the wrong timing. Where you have a choice, record in a common format (mp3, wav, m4a, ogg or webm) at a standard sample rate, and avoid passing the file through multiple apps before transcribing it.

Choose the right model for the job

Tiny is fast and fine for clear English. Base handles accents and overlapping voices better. Small is the most accurate and the slowest, and it is the right choice when a name or a number must be right. Multilingual variants transcribe non-English speech; English-only variants are faster and more accurate on English audio. Start with Tiny, and move up only when the transcript needs it.

Check the result against the audio

No transcript is perfect, especially with accents, crosstalk or poor audio. Read the transcript once while listening to the recording, at least for the parts that matter โ€” names, numbers, dates and decisions. The model renders sounds into text; it does not know whether it heard "Acme" or "a team," so the important words deserve a second check.

Crosstalk is the most common failure mode

Interruptions, overlapping voices and people finishing each other's sentences are hard for any transcription model. Whisper has to separate two sound patterns that occupy the same slice of time, and it cannot always. Meetings with a clear turn-taking habit produce cleaner transcripts than meetings where several voices speak at once.

Accents and uneven training data

A voice the model has heard rarely in its training data is a voice it guesses about more. That is not a judgment about the speaker; it is a statement about the training data. If you know a speaker's accent is under-represented in speech corpora, expect more errors for their utterances and check proper nouns and numbers especially carefully.

Model size changes speed more than quality for clear speech

Tiny transcribes fast and is usually good enough for clear English at a steady pace. Base and Small cost more time and battery; Small is worth it when a proper name or a figure must be exact. The gain from a bigger model is smaller than the gain from moving the microphone closer to the speaker.

Encode once, then transcribe

Every re-encode through a messaging app, a voice recorder or an editor can degrade the audio before it ever reaches the tool. Where you can, record the file once at a standard rate and hand that original to the tool, rather than a copy that has passed through several apps and gained a layer of compression each time.

Transcribe a meeting locally โ€” free and private

← All meeting-notes guides