Methodology
How we test voices and transcription
Every number on our voice pages comes from audio we generated and measured ourselves. This page explains what we measured, with which tools, on which dates — and what the numbers do not tell you.
Last updated 30 October 2026 · transcription and translation measurements added
1. Voice measurements
For each of our 361 voices we generated a sample at normal speed and analysed the audio file. Voices in the same language and engine read the same text, so their numbers can be compared directly. English voices additionally read the same 59-word passage:
Welcome back to the channel. Today we are going to look at three simple ways to make your videos sound better. First, choose a quiet room. Second, keep the microphone close to your mouth. Third, listen to a short test before you record the whole script. Small changes like these make a big difference to every video you publish.
| Measure | How we get it | Measured |
|---|---|---|
| Pitch (median) | Median fundamental frequency of the voiced parts, in hertz. Measured with Praat; for English voices on the 59-word passage, otherwise on the sample sentence. Where Praat gave no value we used an autocorrelation pitch tracker limited to 60–400 Hz. | 16 & 22 Oct 2026 |
| Speaking rate | Words per minute while reading the 59-word passage, timed from the first to the last speech sound (silence at both ends removed). English voices only. | 16 Oct 2026 |
| Intonation range | Distance in semitones between the 10th and 90th percentile of the pitch track — a wider range sounds more expressive, a narrower one steadier. | 22 Oct 2026 |
| Brightness | Spectral centroid of the speech, in hertz. Higher values sound brighter or sharper, lower values warmer. | 22 Oct 2026 |
| Read time | Seconds needed for the shared sample sentence. Only compared between voices that read the same sentence. Replaced on voice pages by the characters-per-second rate below. | 22 Oct 2026 |
| Characters per second | Every voice of a language reads the same ~15-second passage (the sample you hear on its page). We divide the passage length in characters, spaces included, by the time from the first to the last speech sound. Voice pages turn this into an estimate of how long 1,000 characters take; the median per language is listed on our voices guide. | 30 Oct 2026 |
| Pitch line | Pitch tracked with Praat every 10 ms over the same passage. Octave jumps are folded back toward the voice’s median, voiced runs shorter than 50 ms are dropped, a 5-frame median filter smooths the track, and the passage is split into 60 equal time slices. Each point is a slice’s median, in semitones from the voice’s median pitch. | 30 Oct 2026 |
Audio was decoded to 16 kHz mono and analysed in 40 ms windows with a 10 ms step. Voice pages show these numbers together with comparisons against other voices in the same language, and our voice picks by use are chosen from them rather than from opinion.
2. Multilingual voice test
On 16 October 2026 we had 12 multilingual voices read the same sentence in 8 languages, plus control voices. Each recording was transcribed by Whisper large-v3-turbo with automatic language detection, and we scored how closely the transcript matched the original text. The voice-by-language results are on the multilingual voices page.
3. When a voice does not pass, we remove it
On 26 October 2026 we transcribed the samples of our locally run (Studio) voices with Whisper and compared the transcript with the text the voice was asked to read. Studio voices for Chinese had 45–82% character errors and Japanese 41–49%, while Microsoft neural voices for the same languages had 0–2%. We retired those 13 Chinese and Japanese Studio voices; anyone who links to one now reaches the closest Microsoft voice instead. Studio voices for Hindi, Spanish, Portuguese, Italian and French scored 0–9%, in line with the Microsoft voices, and stayed.
4. Subtitle timing
Paid plans get SRT and VTT subtitles with every voiceover. In October 2026 we checked the subtitle start times against word timings from Whisper on the same audio: the average offset was 26–97 milliseconds depending on the voice engine.
5. Transcription and translation accuracy
Most files are transcribed with Whisper large-v3-turbo, which recognises 99 languages; if that service is unavailable, a backup engine finishes the job. On 30 October 2026 we measured it language by language:
- Recordings. The first 20 distinct sentences of the test set of Google FLEURS (CC BY 4.0) for each language — native speakers reading news-style sentences, with the correct text known.
- Same setup as the product. Each clip went through the same Whisper large-v3-turbo call our transcriber makes, with the language selected, as on our language pages.
- Scoring. Both texts are normalised (Unicode NFC, lower case, punctuation removed). Word error rate counts substituted, missing and extra words; character error rate does the same per character, ignoring spaces. Languages written without spaces between words (Chinese, Japanese, Thai, Burmese) are scored by characters only.
- What pages show. Instead of percentages, each language gets a plain rating of how much editing to expect — quick read-through, light, moderate or heavy editing, or not usable yet — based on its typical (median) sentence, so one sentence where the model repeats or invents text does not decide the rating. Pages also show a real example sentence with the wrong words underlined.
- Translation. FLEURS sentences have professional English translations. We translated our transcripts with the same AI step the product uses and compared the result with the English reference using chrF (a standard character-level match score). We also translated the exact source text the same way, so the difference between the two shows how much quality is lost to speech recognition.
- Speed. Waiting times come from real customer jobs: the time between a finished upload and a ready transcript, by recording length.
Read speech in a quiet room is an easier case than interviews, phone calls or recordings with music, so treat these figures as the best case. Results per language are on each audio-to-text page and the audio translator. Languages where most of the text came out wrong are marked "not usable yet", and we do not publish guides that recommend them.
6. What these numbers do not tell you
- They describe our own generated samples, not how a voice will sound with your script. Long texts, names and mixed languages can change the result.
- They are acoustic measurements, not listener ratings. A lower pitch is not “better”; it is lower.
- Pitch trackers can miss very low or very breathy voices; a few voices therefore have fewer measures than others.
- Microsoft and the open-source models we use can update a voice at any time. The dates above tell you when each measure was taken.
7. Corrections
If you find a number on this site that looks wrong, tell us on the contact page or at audioflowteam@speakscribe.me. We re-measure and correct the page, and update the date above.