Model reference — Audio
Noise suppression removes what is behind the voice — traffic, fans, keyboards, room echo, crowd noise — and speech enhancement goes further, restoring a recording that was quiet, distorted or clipped so it becomes clear.
There are two reasons to do it and they pull in different directions. Cleaning audio for a person to listen to is judged by how pleasant it sounds. Cleaning audio for a transcription model is judged only by whether the transcript improves — and aggressive cleaning can make transcription worse, by removing parts of the speech signal along with the noise. So enhancement is always measured against the actual downstream purpose, not on how clean the result sounds.
Models in this group take an audio file as input: a recording, a call, an interview or a broadcast. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one hour of 16 kHz mono speech.
Vendor: Friedrich-Alexander-Universität
What it does: removes steady background noise in real time with very little computing power, small enough to run alongside a live call on ordinary hardware. The usual default for live audio.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 4 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 8 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours processed per hour) | ≈ 200 | ≈ 700 | ≈ 1,800 |
Vendor: Xiph.Org Foundation
What it does: a classical noise suppressor needing no GPU at all. Less effective than the model-based options on difficult noise, but effectively free and dependable on a fan or hum.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | No GPU required | No GPU required | No GPU required |
| VRAM | — | — | — |
| vCPUs | 2 | 4 | 16 |
| RAM | 4 GB | 8 GB | 32 GB |
| Server | CPU instance, 2 vCPU | CPU instance, 4 vCPU | CPU instance, 16 vCPU |
| Rate (audio hours processed per hour) | ≈ 400 | ≈ 1,200 | ≈ 4,000 |
Vendor: Resemble AI
What it does: repairs badly damaged recordings — quiet, distorted, echoing — and rebuilds speech detail rather than merely removing noise. The choice for archive material that must be made listenable.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours processed per hour) | ≈ 8 | ≈ 30 | ≈ 80 |
Vendor: SpeechBrain
What it does: enhances speech optimising directly for perceived quality, which makes it a good fit where a person will listen to the result rather than a machine.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours processed per hour) | ≈ 40 | ≈ 150 | ≈ 400 |
Vendor: Meta
What it does: separates the recording into speech and everything else, then keeps the speech. Heavier than a noise suppressor but far better where the interference is music or another voice rather than steady noise.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (audio hours processed per hour) | ≈ 6 | ≈ 24 | ≈ 60 |
Which model helps depends on the kind of noise, the recording channel and what the audio is for. Our consultants test the candidates on your own recordings, measuring listening quality or transcription accuracy as appropriate, and recommend the one that improves your outcome rather than the one that sounds cleanest.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for Noise suppression / speech enhancement. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.