What is SNR?
Signal-to-Noise Ratio, how much louder the speech is than the background noise, measured in decibels (dB).
Higher is cleaner. Studio recordings usually reach 30 dB or more; a busy restaurant may sit at 5 dB.
When buying data for a noise-robust model, describing it as noisy is meaningless. The SNR band and the noise type have to be specified, or the data cannot be reproduced.
Types of SNR bands
SNR is usually discussed in bands rather than as an exact figure, because a single recording drifts across the range as the speaker turns, moves or pauses. At the top sit studio and close-microphone recordings, where the speech is far above the noise floor and the room is effectively silent. Below that is the ordinary indoor band: an office, a home, a car with the engine running.
Below that is the band where most difficult speech data lives: streets, cafes, factories and kitchens, where noise competes with the voice and recognition accuracy drops. The bottom of the range is speech that is audible but uncomfortable, the busy restaurant where a listener has to lean in. Reverberation travels with it: a large quiet hall can have a good ratio and still be hard.
Noise type matters as much as level. Steady noise, such as a fan, an air conditioner or an engine hum, holds still, and a model can learn to subtract something that does not change. Transient noise, such as a door, cutlery or a cough, arrives without warning. Competing speech is hardest of all, because another voice has the same structure as the one being recognised.
What to specify when buying
The band, first. Describing a set as noisy tells a buyer nothing, because a cafe and a factory are different problems. State the target range, the tolerance around it, and how the figure is measured, whether over the speech segments or across the whole file, which give different numbers. Then say whether the range applies to every file or only to the average.
Then the noise itself, named, and whether it was recorded with the speech or added afterwards. Real noise captured alongside the voice is natural and hard to reproduce. Mixed-in noise is controllable: one clean recording plus a noise source can generate the same speech at several ratios, which is why augmented sets cost less than in-the-wild ones.
Finally the recording setup, because it is what actually produces the ratio: microphone type, distance from the mouth, and the room. A close-talk headset and a microphone across the table yield entirely different data from the same conversation. And ask for a sample with the noise left in. A supplier who denoises before delivery has removed the thing a noise-robust model is trained on.
Related terms
-
Noisy Speech
Speech collected under background noise (street, in-car, restaurant, office and other real environments), used for noise-robust models and speech enhancement.
-
In-the-Wild Speech
Speech collected under fully natural conditions (no studio, no topic constraints), as close to real usage as collection gets.
Keep reading
-
Full glossary
Every term we explain, in one list.
-
How to buy training data
Where these terms actually show up, and which ones change a quote.
-
Help center
How a project runs, from specification to delivery.
Submit a sourcing request
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.