Why speaker count matters more than hours

Ask ten vendors for a quote and nine will answer in hours. That is because hours are easy to count, easy to compare, and easy to inflate. It is also why so many buyers end up with datasets that look adequate on paper and underperform in training.

What a model actually learns from

When a speech model trains on a dataset, it is not learning "English" or "Arabic." It is learning the acoustic patterns present in the recordings you gave it: including the individual characteristics of the people who happened to be recorded.

If those recordings come from a small group of speakers, the model learns that group well and everything else poorly. This is speaker overfitting, and it is invisible in a cost-per-hour comparison.

A hundred hours from five speakers is, for generalization purposes, close to five hours of distinct information repeated twenty times.

What to ask for instead

Ask for the speaker manifest. Any serious producer keeps one, because they need it to manage the recording schedule.

• How many distinct speakers, and what is the minimum per speaker?
• What is the demographic spread: age range, gender balance, region or dialect background?
• How were speakers recruited, and does that recruitment method bias the pool?
• Is any speaker represented disproportionately in the total hours?

Why vendors resist this framing

Speaker count is expensive to increase in a way that hours are not. Recording the same five people for another twenty hours costs studio time. Recruiting, screening, and scheduling fifty new speakers costs coordination, and for uncommon languages it can cost weeks of lead time.

That is a real cost difference, and it is why a per-hour price is a poor comparison across vendors: the cheap quote and the expensive one may not be selling the same thing.

The exception

There is one case where hours genuinely dominate: single-speaker voice cloning and personalized TTS, where the goal is to model one specific voice as accurately as possible.

That is a narrow case. For almost every other application (recognition, diarization, wake words, emotion classification), the speaker count is the number that predicts whether the model works outside the lab.

Writing the requirement so it survives procurement

A speaker count buried in a paragraph of prose will not survive contact with a purchasing department that compares quotes by the hour. Put it in the specification as a numbered, testable line, and put it in the contract as an acceptance criterion. Four lines do most of the work:

• Minimum distinct speakers, stated as a floor rather than a target. "At least 400 distinct speakers" is enforceable; "approximately 400" is not.
• Maximum hours per speaker. This is the line that actually caps the concentration. Without it, a producer can meet the speaker floor and still fill half the delivery from the ten most available people.
• Recruitment method disclosure. How the pool was assembled determines what it represents, and it is the first thing you cannot reconstruct after the fact.
• A speaker manifest as a delivery artifact: one row per speaker, with an opaque speaker ID, session count, and total duration. Not names. You need the distribution, not the identities.

The counter-argument you will hear, and the answer

Someone in the room will point out that the per-hour price is lower at the vendor who records fewer people for longer, and that they are right about the arithmetic. They are. The number is lower. What it buys is not the same product, and the argument for paying more is not about quality in the abstract.

The framing that tends to land is the cost of finding out. A model trained on a concentrated corpus fails in production, and by the time it fails you have already paid for the training run, the evaluation cycle, and the deployment. The dataset is the cheap part of that sequence, which makes a re-collection the expensive way to save money on it.

The other half of the answer is how you test the claim. Hold out speakers, not utterances. A random utterance split puts the same voices in both the training and the evaluation set, and the resulting score measures memorization rather than generalization. A speaker-disjoint split is the only one that answers the question you actually asked.