Practical guides
How much audio do you need to fine-tune an ASR model?
How many hours of speech data you actually need to fine-tune an ASR model, with realistic thresholds for 200, 500, 2,000 and 10,000 hours.
Key takeaways
- 200–500 hours is where fine-tuning a multilingual ASR model starts producing gains you can measure in production, not just on a held-out test set.
- 1,000–3,000 hours is the range where most teams stop seeing dramatic returns and start seeing incremental ones.
- Distribution beats volume. 500 hours matched to your deployment conditions outperforms 5,000 hours of mismatched audio, consistently.
- Training from scratch needs tens of thousands of hours and is almost never the right call for a low-resource language.
The short answer
For fine-tuning an existing multilingual model such as Whisper, MMS or a wav2vec 2.0 checkpoint, meaningful gains typically begin around 200 to 500 hours of well-matched, accurately transcribed speech in your target language. Below roughly 100 hours you will usually see movement on a held-out test set and disappointingly little in production.
That range assumes the audio actually resembles what your model will encounter at inference. If it does not, no amount of volume rescues you — which is the single most common and most expensive mistake teams make.
Why the number depends entirely on the task
A constrained-grammar task — digit strings, addresses, a fixed command set — can reach production quality on 50 to 150 hours, because the search space is small and the model mostly needs to learn the acoustics of your speakers.
Open-domain conversational ASR is a different problem. Spontaneous speech carries disfluency, overlap, repair, laughter, filler and code-switching, none of which appear in read speech. Budget 1,000 hours or more, and expect the first 300 to deliver most of the improvement.
Telephony is harder again. An 8 kHz narrowband channel discards most of the spectral information a wideband model was trained on, so telephony audio is effectively a separate domain requiring its own data rather than a subset of your existing corpus.
Distribution matters more than volume
This is the part most guides skip. A model can only learn the world its data describes. If your 500 hours come from young, urban, university-educated speakers on flagship handsets, your model will be excellent for young urban graduates with good phones and mediocre for everyone else.
We have repeatedly seen 500 well-distributed hours beat several thousand convenience-sampled hours on real production traffic. The mechanism is simple: convenience sampling over-represents whoever was easiest to recruit, and the tail — where churn actually lives — stays unlearned. This is why our collection is quota-controlled rather than volume-driven.
A practical budgeting table
As rough planning figures from completed projects: 50–150 hours for wake word or command recognition. 200–500 hours for meaningful fine-tuning gains on a multilingual base. 1,000–3,000 hours for competitive open-domain conversational ASR in one language. 20–60 hours of studio audio per voice for TTS, which is a different task with different requirements entirely.
Training a foundation ASR model from scratch is an order of magnitude beyond all of this — tens of thousands of hours — and for a low-resource language it is almost always more efficient to fine-tune a strong multilingual checkpoint.
How to spend the first 200 hours well
Do not order 2,000 hours on day one. Commission a small, deliberately diverse pilot — 50 to 200 hours — that spans your full quota spec rather than a slice of it. Fine-tune, measure, and look specifically at which speaker segments improved and which did not.
That result tells you where the remaining budget should go. A pilot that costs a fraction of a production order routinely prevents a re-collection that would have cost far more, which is why we credit pilot fees against production work.
Answers
Frequently asked questions
For a language Whisper already covers weakly, 200–500 hours of well-matched, accurately transcribed audio usually produces clearly measurable Word Error Rate improvement. For a language it handles reasonably already, you may see gains from as little as 50–100 hours of in-domain audio that matches your acoustic conditions. Matching the deployment condition matters more than raw hours.
No. Additional data that duplicates the distribution you already have produces diminishing and eventually negligible returns. Additional data that fills a gap — a dialect, an age band, a device class, a noise condition you were missing — produces disproportionate returns. Audit what your corpus lacks before buying more of what it already has.
As augmentation, sometimes usefully. As a replacement for real speech, no. Synthetic audio lacks the disfluency, channel variation, overlap and background structure of real recordings, and models trained heavily on it tend to fail in exactly the conditions you deployed them for.
Not sure how many hours your model needs?
Describe the failure you are seeing. We will tell you what collection design fixes it and roughly how much data it takes — including when the answer is less than you expected.
Average first response: under 6 business hours. NDAs signed same day.