Model reference — Audio
Source separation splits a mixed recording into its parts — the voice on one track, the music on another, drums and bass on their own — so that each can be handled separately.
The uses are practical. A broadcast can be transcribed accurately once the background music is removed. A film’s dialogue can be re-recorded in another language while the score and effects are kept. An interview recorded in a bar becomes usable. A recording where two people talk over each other can be split so both are transcribed. Separation is never perfect: some artefacts remain, particularly where two sources occupy the same frequencies, which is why the result is judged against its purpose rather than in isolation.
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: Meta
What it does: separates a recording into voice, drums, bass and other instruments with the best quality generally available. The default for broadcast, film and music work.
| 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 |
Vendor: Kuielab
What it does: a faster separator that isolates voice from accompaniment with quality close to the best, at a fraction of the cost per hour. The economic choice for large volumes.
| 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) | ≈ 15 | ≈ 55 | ≈ 140 |
Vendor: Resemble AI
What it does: rebuilds a separated voice track that came out thin or artefact-laden, which is often the step that makes separation output good enough to broadcast.
| 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: Friedrich-Alexander-Universität
What it does: a lightweight option that pulls speech out of steady background sound in real time, suited to live use where full separation is too slow.
| 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: pyannote (Hervé Bredin)
What it does: detects where two people are speaking at the same time, which tells a pipeline which stretches need separating rather than separating the whole file.
| 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) | ≈ 20 | ≈ 70 | ≈ 180 |
The right model depends on what you are separating and whether the output is for a listener or a transcription model. Our consultants test the candidates on your own material and recommend the model, the number of tracks and the settings that give the cleanest usable result.
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 Speech/music source separation. 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.