Model reference — Audio
Speaker recognition puts a name to a voice. Verification answers a yes-or-no question — is this the person who enrolled? — and identification searches a list of enrolled voices for the closest match.
Both work by turning a few seconds of speech into a voice fingerprint and comparing it against fingerprints recorded at enrolment. Verification is the easier and more common case, used to confirm a caller’s identity. Identification against a large list is harder, because every additional enrolled voice is another chance of a false match, so the decision threshold must be set against the size of the list. Two cautions apply in every deployment: voices change with illness and age, so enrolment must be refreshed; and recorded or synthesised speech can be presented to the microphone, so a liveness check belongs in any pipeline used for access.
Models in this group take a short speech sample as input and compare it against enrolled voice fingerprints. 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 five-second speech sample compared against an enrolled voice.
Vendor: NVIDIA
What it does: turns a few seconds of speech into a voice fingerprint accurate enough for verification, and is small and quick enough to run on every call. The usual first choice.
| 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 (verifications/hour) | ≈ 60,000 | ≈ 220,000 | ≈ 600,000 |
Vendor: SpeechBrain
What it does: a well-established fingerprinting model with excellent accuracy on telephone-quality audio, which is where most verification actually happens.
| 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 (verifications/hour) | ≈ 70,000 | ≈ 260,000 | ≈ 700,000 |
Vendor: Northwestern Polytechnical University
What it does: a production toolkit covering enrolment, fingerprinting and searching a large enrolled list, which is what identification rather than verification requires.
| 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 (verifications/hour) | ≈ 50,000 | ≈ 190,000 | ≈ 520,000 |
Vendor: pyannote (Hervé Bredin)
What it does: provides fingerprints tied to the same models used for diarization, which keeps one pipeline for both "who spoke when" and "who is this".
| 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 (verifications/hour) | ≈ 40,000 | ≈ 150,000 | ≈ 400,000 |
Vendor: Silero
What it does: detects whether the sample contains real speech at all before verification runs, which removes a whole class of failure caused by judging silence, noise or a tone.
| 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 (verifications/hour) | ≈ 400,000 | ≈ 1,500,000 | ≈ 4,000,000 |
The right configuration depends on how many voices are enrolled, the cost of a false match against a false rejection, and whether the channel is telephone or high-quality audio. Our consultants measure both error rates on your own recordings, then recommend a model, a threshold and the liveness checks the use case requires.
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 Speaker recognition / verification. 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.