How accurate is free AI transcription? We tested it
Tested and published October 31, 2026
"Accurate" is easy to claim and hard to pin down, so we measured it. We took 100 public test recordings with known correct transcripts, ran them through both of FreeTranscribe's settings, and also played the clean recordings back with background noise and squeezed through phone-call quality. Here's what came out, what it means for your recordings, and how to get the best transcript you can.
The results
The number below is the word error rate: out of every 100 words spoken, how many the transcript got wrong, left out, or added. Lower is better. Around 5% means roughly one slip every two sentences; most people find a transcript easy to clean up below 10%.
| Recording | Accurate setting | Fast setting |
|---|---|---|
| Clear speech, quiet room | 5.1% | 16.0% |
| Clear speech with moderate background noise | 7.9% | 95%* |
| Phone-call quality | 9.8% | 37%* |
| Harder speakers (accents, faster or less clear speech) | 12.4% | 38.0% |
| Loud background noise (as loud as the voice) | 25.6% | 239%* |
* Rates above 100% mean the model added more text than was spoken. On noisy audio the Fast setting sometimes got stuck repeating a phrase ("the matter is that the matter is that…") for the rest of the clip: 3 of 60 clips with moderate noise, 9 of 60 with loud noise, and 1 of 60 phone-quality clips. Leaving those clips out, Fast scored about 26% with moderate noise and 30% on phone audio. The Accurate setting never did this in our test.
What it sounds like in practice
Here's one sentence from the test, exactly as each setting wrote it:
| What was said | Fortunately, said Mr. Vanderpool, northerners and southerners are arriving at a better mutual understanding on most of these matters. |
|---|---|
| Accurate, clean | Fortunately, said Mr. Vanderpool, northerners and southerners are arriving at a better mutual understanding on most of these matters. (perfect) |
| Fast, clean | Fortunately, said Mr. Vanderpool, Northerners and Sothers are arriving at a mutual understanding on most of these matters. (dropped "better", misheard "southerners") |
| Accurate, phone quality | Fortunately, said Mr. Vanderfool, Northern of the Southerners are arriving at a better mutual understanding on most of these matters. |
| Accurate, loud noise | Fortunately, still can still be done with the Northerners and Southerners, all arriving at a better mutual understanding on most of these matters. |
Notice the pattern: the errors aren't random letters, they're plausible words. A name becomes a similar-sounding word, a short phrase gets swapped for something that sounds alike. That's why a quick read-through catches most of them.
What this means for you
- Use Accurate unless you have a reason not to. It made about a third as many mistakes on clean speech and held up far better on difficult audio. Fast is there for older phones and laptops and for quick drafts of clear recordings.
- Moderate background noise barely matters; loud noise does. Steady noise well below the voice (10 dB quieter in our test, like a fan or a quiet office) took Accurate from 5% to about 8%. Noise as loud as the speaker, closer to a busy café or music under the voice, made it about five times worse.
- Phone calls and voice notes roughly double the errors compared with a decent microphone, because the call strips out the high frequencies that make consonants clear.
- If a Fast transcript starts repeating itself, run it again on Accurate. Repetition loops were the Fast setting's biggest weakness in our test.
How to get a more accurate transcript
Before you record
- Get the microphone close to the speaker. Distance makes everything harder for the model, more than people expect.
- Avoid music or a TV under the voice. Our loud-noise result is the reason.
- For calls and interviews, record each side locally if your app allows it, rather than the call audio.
When you transcribe
- Choose Accurate.
- Set the spoken language yourself. Auto-detect is usually right, but a wrong guess on a quiet intro or a short clip hurts the whole transcript.
- If the recording is mostly music, intros or silence, trim those first; the model can invent words in long silent or musical stretches.
When you check it
- Read names, places, brands and technical terms first; that's where most mistakes are.
- Watch for small swapped words in negatives and numbers ("can" vs "can't", "fifteen" vs "fifty"), which change meaning.
- Use the Timestamps tab to find the spot in the recording when something looks off.
- Numbers are written as digits ("August 27th, 1837"), which is usually what you want.
How we tested
- Recordings: 60 utterances from 40 different speakers in the LibriSpeech "test-clean" set (8.3 minutes of speech, 1,321 words) and 40 utterances from 32 speakers in the harder "test-other" set (3.8 minutes, 611 words), spread evenly through each set. LibriSpeech is a widely used public speech-recognition benchmark built from public-domain audiobooks (openslr.org/12, Panayotov et al., 2015).
- Conditions: the clean recordings as-is; with pink noise mixed in at 10 dB and at 0 dB below the speech (moderate and loud); and band-limited to 300–3,400 Hz with 8-bit companding, which approximates a phone call.
- Models: exactly what the site runs: Whisper base (Accurate) and Whisper tiny (Fast), the same 8-bit ONNX versions and settings, with the language set to English. We ran them on a desktop computer outside the browser; the browser runs the same model files, so its results should match closely.
- Scoring: word error rate after normalising capitals, punctuation and how numbers are written (so "Mr." matches "mister" and "1837" matches "eighteen thirty seven").
- Limits: this is a small, English-only test on read speech; real conversations with crosstalk are harder. For comparison, OpenAI's published results for Whisper base on LibriSpeech test-clean are also about 5%, so the browser version matches the original. The compressed tiny model loses noticeably more than the original tiny model does.