Model reference — Audio
Diarization works out who spoke when. It divides a recording into stretches and labels each with a speaker — Speaker 1, Speaker 2 — without knowing who those people are, so that a transcript reads as a conversation instead of an undivided block of text.
It is what makes a meeting or call transcript usable: you can see how long each person talked, follow an exchange, and attribute a statement to the person who made it. The hard cases are overlapping speech, where two people talk at once, and short interjections of a second or two. Diarization answers "how many speakers and when"; putting real names to those speakers is speaker recognition, a separate service that needs enrolled voice samples.
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: pyannote (Hervé Bredin)
What it does: divides a recording by speaker and handles overlapping speech, which is where most diarization fails. The most widely deployed option and the usual default for meetings and interviews.
| 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 |
Vendor: NVIDIA
What it does: a newer approach that labels speakers in a single pass rather than clustering afterwards, which improves accuracy on recordings with many participants and frequent turn-taking.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 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) | ≈ 25 | ≈ 90 | ≈ 220 |
Vendor: NVIDIA
What it does: produces a voice fingerprint for each stretch of speech, which is the component that decides whether two stretches are the same person. Small and fast.
| 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) | ≈ 80 | ≈ 300 | ≈ 800 |
Vendor: SpeechBrain
What it does: an alternative voice-fingerprinting model, well established and easy to fit to a specific recording setup such as a particular conference room or telephone line.
| 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) | ≈ 90 | ≈ 340 | ≈ 900 |
Vendor: Northwestern Polytechnical University
What it does: a production-oriented fingerprinting and clustering toolkit, suited to running diarization over large call archives at a predictable cost.
| 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) | ≈ 70 | ≈ 260 | ≈ 700 |
Vendor: University of Oxford (VGG)
What it does: aligns a transcript to the audio to the word and attaches speaker labels to it, which is what turns separate transcription and diarization outputs into one readable, timed transcript.
| 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 |
Accuracy depends on how many speakers there are, how much they overlap, and whether the recording is one microphone or several. Our consultants measure diarization error on your own recordings, then recommend a model and a pipeline that fits your call or meeting format.
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 diarization. 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.