How do you transcribe low-quality or quiet audio?

Fix the recording first, then transcribe it, then check the parts you cannot hear. Background noise and distance from the microphone reduce accuracy, so start with those rather than with the software. Keep the original file rather than a compressed copy, normalize the loudness of a copy so quiet speech is easier to check, and listen closely to every silent or near-silent passage. The guide to transcribing audio to text covers the full process, and the transcription overview links the other guides.

Noise costs accuracy, so fix the source before you choose software

Noise is a common reason a transcript fails. A 1996 study found that the performance of a speech recognition system degrades rapidly in the presence of ambient noise ¹. A 2017 study of speech recognition in conversation said that accuracy for any application can be affected by speech fidelity, ambient noise and microphone quality ². You can improve the ambient noise and the microphone at the recording stage.

Distance is the cheapest fix. Under ideal free-field conditions, the inverse-square rule says a microphone at 12 inches receives about 21.6 dB less level than one at 1 inch ³. A real room has echo and reflections, so the gain in practice is smaller, but the direction holds: get the microphone close to the speaker. Apple gives similar advice. It says a headset microphone may help when dictation is used in a noisy environment or a room with excessive echo ⁴.

If you can record again, a noise-canceling microphone reduces the room sound that reaches the recorder. For a recording you already have, the steps below are what you can do.

Compressed copies lose detail, so transcribe the original

Google's API reference says the accuracy of its speech recognition can be reduced if lossy codecs are used to capture or transmit audio, particularly if background noise is present ⁵. The same passage lists MP3 and OGG_OPUS among the lossy codecs.

Two rules follow from that warning. Keep the original file from your recorder, and transcribe that rather than a copy that has been converted several times. If you must convert, use the highest quality setting that the format allows, and keep the converted file apart from the original.

Normalizing the level makes quiet speech easier to hear, but it does not remove noise

The loudnorm filter in ffmpeg performs EBU R128 loudness normalization, in both dynamic and linear modes ⁶. In practice it raises or lowers the whole file toward a target loudness, so quiet speech becomes audible when you check it. It changes how loud the recording is. It does not remove hiss, hum or room noise, so it does not replace a better recording.

This is a standard ffmpeg command. Run it on a copy and test it on a short clip first:

ffmpeg -i quiet-interview.m4a -af loudnorm -ac 1 normalized.wav

Then transcribe both the original and the normalized copy, and keep whichever gives the more accurate text on the parts you can hear clearly.

Quiet passages can produce invented text, so check them by ear

Silence and near-silence are where a recognizer is most likely to invent words. In a 2024 study of a 2023 version of Whisper, on recordings from people with and without aphasia, roughly 1 percent of transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ⁷. The usual sign is a phrase that repeats, or text that appears over a gap where nobody speaks.

If you run open-source Whisper yourself, its transcribe function has a no_speech_threshold setting. A segment counts as silent when its no-speech probability is above that value and its average log probability is below the logprob_threshold ⁸. Changing the threshold changes which passages are dropped. It can also drop quiet speech that is real, so check the output after you change it.

How to transcribe a quiet recording, step by step

1. Listen first. Use headphones, turn the volume up, and note every passage you cannot hear clearly.
2. Record again if you can. Move closer to the speaker, or use a headset microphone in a noisy room.
3. Normalize a copy. Run the loudness command on a copy and keep the original untouched.
4. Transcribe and compare. Transcribe the original and the normalized copy, and keep the one that is more accurate on the clear parts.
5. Check the quiet parts. Look for repeated phrases and text that appears over silence, and delete any passage that was never spoken.
6. Mark what you cannot hear. Write [inaudible] rather than guessing, especially in a formal record.
7. Decide whether a person should review it. A low-quality recording with a high cost of error may need a human transcriptionist. The comparison of human and AI transcription covers the trade-off.

Accuracy is measured as the word error rate, which is the share of words a tool gets wrong. A rate measured on clean speech will not apply to your noisy recording, so judge the result by listening to it.

For the formats a tool accepts, see which audio formats can be transcribed. For the privacy side of the choice, see how to transcribe audio offline.