Model reference — Audio
Audio semantic search finds the moment in a recording that matches what you asked for, whether that is something said — "where did they discuss the Bristol contract?" — or something heard — "find the alarm".
Two indexes serve the two kinds of question. For speech, the recording is transcribed and the transcript is indexed by meaning, so the search behaves like document search and returns a timestamp. For sound, the audio itself is fingerprinted against written descriptions, which finds non-speech events that no transcript would contain. Most archives want both, built in the same pass over the audio, since the expensive part is listening to the file once.
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: SYSTRAN
What it does: transcribes the archive with timings so speech becomes searchable text pointing back to the second it was said. The fastest practical way to index a large volume of speech.
| 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) | ≈ 20 | ≈ 70 | ≈ 180 |
Vendor: OpenAI
What it does: the more accurate transcription option for archives where a missed word means a missed search result, at roughly a fifth of the speed.
| 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) | ≈ 4 | ≈ 12 | ≈ 30 |
Vendor: Microsoft and LAION
What it does: fingerprints the audio itself so it can be searched with a written description of a sound, which finds events that appear in no transcript.
| 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) | ≈ 70 | ≈ 260 | ≈ 700 |
Vendor: Microsoft
What it does: labels sound events across several hundred categories, giving a structured index — every alarm, every stretch of music — alongside the free-text search.
| 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) | ≈ 60 | ≈ 220 | ≈ 550 |
Vendor: Beijing Academy of Artificial Intelligence
What it does: indexes the transcripts by meaning, so a search finds the passage that answers the question rather than only the words that were typed.
| 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) | ≈ 6,000 pages/hour | ≈ 30,000 pages/hour | ≈ 90,000 pages/hour |
Vendor: Beijing Academy of Artificial Intelligence
What it does: re-sorts the candidate moments by how well they actually match the question, which is what puts the right moment first rather than fifth.
| 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 (query-passage pairs/hour) | ≈ 40,000 | ≈ 150,000 | ≈ 400,000 |
The right pipeline depends on whether your questions are about words, sounds or both, and how large the archive is. Our consultants review your audio and the searches you need, then recommend the transcription and fingerprinting models and how the index should be built.
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 Audio semantic search. 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.