Voice Data Sets by the Numbers: What Public Speech Corpora Actually Contain

When you are choosing a voice data set to train or evaluate a speech model, the headline that matters is not "how big" in the abstract. It is how many hours, how many speakers, how many distinct utterances, and how that supply is shaped. A corpus built for broad multilingual coverage behaves very differently from one built to recognize a short list of spoken commands, and a single-speaker audiobook set is a different animal again. The public figures below show just how far these corpora spread, and they are a useful reference point for anyone scoping a voice collection of their own.

Broad multilingual speech is measured in thousands of hours

The clearest signal of a general-purpose voice corpus is raw recorded time across many languages. Mozilla's Common Voice set holds 7,335 validated hours of speech in 60 languages (Stanford CS224S dataset list). That shape, a lot of validated audio spread across dozens of languages, is what teams reach for when they need coverage rather than depth in any single accent or domain. The breadth is the point: sixty languages in one set is a different procurement problem than one language collected deeply.

Command and intent sets are large in count but narrow in vocabulary

Keyword and intent data looks big by utterance count while staying deliberately small in vocabulary. Google Speech Commands contains 65,000 one-second long utterances of 30 short words, recorded by thousands of different people (Stanford CS224S dataset list). Fluent Speech Commands takes a tighter approach, with 30,043 utterances from 97 speakers. Put those two side by side and the trade-off is obvious. One leans on many voices saying a handful of words, the other on fewer speakers producing more varied phrasing. Both are command-focused, and neither is trying to be a general speech corpus.

Labeled audio events reach into the millions

When the task moves from transcribed speech to recognizing sounds, the labeling scale jumps. AudioSet covers 632 audio event classes and a collection of 2,084,320 human-labeled 10-second sound clips drawn from YouTube videos (Stanford CS224S dataset list). Two million human-labeled clips is a reminder that audio data sets are not only about words. The human labeling behind a number like that is the expensive part, and it is why sourcing this kind of data at quality is a specialty rather than a scrape.

That expert layer is where Bunny Labs focuses, delivering expert-driven data collection, evals, and RL environments for large language models, world models, robotics, and autonomous driving, backed by a network of 3 million+ professionals.

Single-speaker TTS corpora trade breadth for consistency

Text-to-speech training inverts the priorities above. Instead of many voices, it wants one voice recorded cleanly and consistently. A representative single-speaker English audiobook dataset holds 13,100 short clips (Codesota TTS benchmark lineage). A number that small, next to a two-million-clip event set, is not a weakness. For TTS, a tightly controlled single speaker is exactly the asset you want, and the count reflects a narrow, high-consistency goal rather than coverage.

Takeaway

There is no single "right size" for a voice data set. The useful questions are what the data is for and how it is shaped. Multilingual recognition points you toward thousands of validated hours across many languages. Command recognition points you toward tens of thousands of short, tightly scoped utterances. Audio-event work points you toward millions of human-labeled clips. Single-speaker TTS points you toward a few thousand clean clips of one voice. Match the shape to the task before you worry about the headline count.

Sources