A machine may transcribe a read sentence without handling an improvised response well. The Common Voice release announced by Mozilla on 18 September 2026 makes this distinction visible: it publishes recordings of prepared speech and spontaneous speech separately. So, to assess speech recognition, counting hours of audio is not enough; we also need to know how people speak.
Two ways of speaking up
Version 27 of Scripted Speech covers 295 languages and around 42 593 hours of recordings, according to the project announcement. Version 5 of Spontaneous Speech brings together free-form responses to questions in 80 languages, with 302 hours transcribed and validated.
These figures describe resources available for developing speech technologies. They are not a ranking of software or a success rate achieved by an application. The distinction between reading and free-form responses mainly serves to identify two use cases that are not equivalent.
A prepared dictation, a hesitant conversation and an interview do not make exactly the same demands on a system. The choice of recordings used to evaluate it is therefore part of the result.
Counting what is missing, changed or added
The word error rate, often called WER, compares an automatic transcription with a human reference text. In the method described by NIST for OpenASR21, deleted, inserted and substituted words are added together, and the sum is then divided by the number of words in the reference.
Calculation example, not a software test: in a reference of one hundred words, five substitutions, three omissions and two insertions make ten errors, or 10%. This figure does not mean that 10% of the sentences are entirely wrong. Nor does it, by itself, show whether an error concerns a detail or a name essential to understanding.
The protocol matters just as much as the formula. OpenASR21 kept training, development and evaluation data separate. Some tracks took capitalisation into account; others did not. Scores published without these details may therefore appear comparable even though they do not measure exactly the same task.
Diversity is also built by listening
Common Voice relies on contributions and their verification. In the procedure described by Mozilla, one person can read a sentence; another then listens to the clip to check that it matches the text. The project says that differences in accent are accepted and encouraged. It also seeks diversity in recording conditions and locations where contributions are made.
This step distinguishes an erroneous transcription from a different way of pronouncing the same words. Rejecting a voice solely because it sounds different would reduce precisely the diversity being sought.
So, when choosing a transcription tool, the useful question goes beyond its best advertised percentage: has it been evaluated on the language, type of speech and sound conditions that resemble the intended use? A larger corpus opens up possibilities. The quality of a particular tool still needs to be measured using recordings it will actually have to process.
Sources
- Mozilla: Common Voice release of 18 September 2026.
- NIST: OpenASR21 protocol and measurement method.
- Mozilla: contributing and validating recordings.
Sources consulted on 21 October 2026.
