How much does background noise cost a contact center?
Under the stated assumptions, background noise costs a 500,000-minute-a-month contact center roughly $474 to $6,000 a month. The WER inputs are measured; the minutes, repeat-rate and cost-per-minute inputs are assumptions that you should replace with your own.
Under the assumptions below, background noise costs a contact center that handles 500,000 minutes a month roughly $474 to $6,000 a month, and the assumptions, not our measurements, set that range. What we can state from measurement is how far a recogniser's word error rate moves: from 4.74% on clean speech to 6.32% at 10 dB of noise, 11.05% at 0 dB and 24.74% in babble at 5 dB.
This is a worked calculation, not a forecast. Replace the four assumptions in the second table with your own call data and rerun the arithmetic.
What we measured
| Condition (raw audio) | WER | Delta vs clean (pp) |
|---|---|---|
| clean | 4.74% | 0.00 |
| noise_snr10 | 6.32% | +1.58 |
| noise_snr0 | 11.05% | +6.32 |
| babble_snr5 | 24.74% | +20.00 |
Measured 2026-08-14, benchmark.json, faster-whisper base.en, 190 reference words per condition, so one word is 0.53 pp. A production recogniser will have different absolute numbers; the ordering is what carries over.
The assumptions (reader-adjustable)
| ID | Assumption (adjust) | Used in step |
|---|---|---|
| A1 | 500,000 | 1 |
| A2 | 20% | 2 |
| A3 | 0.5% | 3 |
| A4 | $0.60 | 4 |
Every row is an assumption, not a measurement. Step 1: monthly minutes handled. Step 2: the share of those minutes that are noisy. Step 3: extra handled minutes, as a share of noisy minutes, for each point of WER (we have no measurement of this relationship; it is the most uncertain input). Step 4: fully loaded cost of one handled minute.
cost_per_month = minutes * noisy_share * (wer_delta_pp * repeat_rate_per_pp) * cost_per_minute
The result
| Scenario | WER delta (pp) | Noisy minutes / month | Extra handled minutes / month | Cost / month |
|---|---|---|---|---|
| noise_snr10 | 1.58 | 100,000 | 789 | $474 |
| noise_snr0 | 6.32 | 100,000 | 3,158 | $1,895 |
| babble_snr5 | 20.00 | 100,000 | 10,000 | $6,000 |
The range is $474 (10 dB noise) to $6,000 (babble at 5 dB), with $1,895 at 0 dB. Double A2 and the cost doubles; double A3 and it doubles again. Because the model is linear, you can scale the whole table by your own factors.
Where our engine moved the WER
The cost above is the gross cost of noise, not a saving. Whether an intervention recovers any of it is a separate measurement, and ours is mixed.
| Condition | WER raw | WER focus v8_2 | Change (pp) |
|---|---|---|---|
| babble_snr5 | 24.7% | 20.5% | -4.2 |
| competing_snr0 | 84.7% | 74.2% | -10.5 |
| competing_snr10 | 6.3% | 12.1% | +5.8 |
| reverb_snr10 | 13.7% | 30.5% | +16.8 |
Measured 2026-08-14. Pooled over all 19 conditions our v8_2 engine scored 16.59% against 14.32% for the untouched audio, so it is not a default-on component. It lowered WER in the babble and the loudest competing-voice conditions and raised it in others. For scale, our price card puts 500,000 processed minutes at $410 a month (10,000 free, then $0.0010, $0.0008 above 100,000), but a price without a measured recovery on your own audio is not a business case.
FAQ
Does WER translate directly into cost? No. The link is assumption A3, and we have no measurement of it. Measure repeat calls or escalations against noise level in your own data.
Which recogniser were these WER values measured with? faster-whisper base.en, int8 on CPU. Other recognisers will give different absolute values.
Will an isolation engine recover this cost? Not across the board. Pooled over our 19 conditions it did worse than raw audio; it helped in specific conditions listed above.