Model reference — Audio
Language detection listens to a few seconds of speech and reports which language is being spoken, so that a recording can be routed to the right transcription model, the right translation pair or the right agent.
It is a small model doing a cheap job that saves a large one from being wasted. Detection is usually reliable within three to five seconds of clear speech, less so on short utterances, heavy accents and closely related languages — Danish and Norwegian, Hindi and Urdu, Malay and Indonesian are the classic confusions. Where a recording contains more than one language, detection is run in windows across the file so each stretch is labelled separately rather than the whole file being forced into one answer.
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 speech labelled in five-second windows.
Vendor: SpeechBrain
What it does: identifies 107 languages from a few seconds of speech. Small, fast and accurate, and the usual first choice for routing.
| 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) | ≈ 200 | ≈ 800 | ≈ 2,000 |
Vendor: OpenAI
What it does: uses the transcription model’s own language detection, which is more accurate on short and noisy clips because it is doing the harder job of understanding the words. Slower, and worth it when a wrong answer is costly.
| 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) | ≈ 40 | ≈ 150 | ≈ 400 |
Vendor: Meta
What it does: identifies over a thousand languages, which is the only practical option for audio arriving in languages outside the common set.
| 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: Silero
What it does: detects where speech is present at all, which is run first so that language detection judges actual speech rather than silence or music. Small enough to run on any hardware.
| 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) | ≈ 600 | ≈ 2,000 | ≈ 6,000 |
Vendor: Meta
What it does: a small model fitted to your own language set, which is more accurate than a general model when you only care about six languages and want the confusions between them minimised.
| 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) | ≈ 150 | ≈ 600 | ≈ 1,600 |
Which model fits depends on your language set, how much speech you have to judge from, and whether recordings switch language mid-way. Our consultants measure detection accuracy on your own audio, then recommend a model, a window length and the confidence threshold below which a recording should be handled as unknown.
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 language detection. 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.