Model reference — Live feed
This service isolates the narrator or presenter from everything else on a live feed — music, effects, crowd noise, other voices — and transcribes only that voice as it is spoken.
The task is separation followed by transcription, both under a delay budget. Separation matters because a commentator over crowd noise or a presenter over a music bed transcribes badly, and cleaning the voice first is usually worth more than choosing a better transcription model. Where several people speak, the narrator is identified by voice, so the transcript follows the one person rather than everyone audible. The whole chain has to fit within a few seconds if the output is captions, and it can be looser if the output is a searchable record.
Models in this group take a live camera or stream as input and must keep pace with it in real time. 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 how many camera streams or feeds one server of that tier can keep up with in real time, not a per-hour count: live work must fit inside the interval between frames, and a server that cannot keep pace drops frames rather than falling behind. 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 1080p stream at 25 frames per second.
Vendor: NVIDIA
What it does: transcribes English in real time at very low delay, which is what makes live captioning possible. The default where speed governs.
| 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 (camera streams handled at 25 frames per second) | ≈ 8 streams | ≈ 30 streams | ≈ 80 streams |
Vendor: SYSTRAN
What it does: transcribes a hundred languages live in short overlapping windows, at a few seconds of delay — right for monitoring and records rather than on-screen captions.
| 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 (camera streams handled at 25 frames per second) | ≈ 3 streams | ≈ 12 streams | ≈ 32 streams |
Vendor: Kuielab
What it does: separates the narrator’s voice from music and effects fast enough to stay live, which is usually the largest single gain in transcription accuracy on a produced feed.
| 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 (camera streams handled at 25 frames per second) | ≈ 2 streams | ≈ 8 streams | ≈ 20 streams |
Vendor: Meta
What it does: the higher-quality separation option, used where the feed is heavily mixed and the extra hardware is justified.
| 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 (camera streams handled at 25 frames per second) | ≈ 1 stream | ≈ 4 streams | ≈ 10 streams |
Vendor: NVIDIA
What it does: identifies which voice is the narrator’s, so the transcript follows that person and ignores others who speak.
| 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 (camera streams handled at 25 frames per second) | ≈ 30 streams | ≈ 100 streams | ≈ 300 streams |
Vendor: pyannote (Hervé Bredin)
What it does: separates speakers live where more than one person presents, keeping their contributions apart in the transcript.
| 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 (camera streams handled at 25 frames per second) | ≈ 4 streams | ≈ 14 streams | ≈ 40 streams |
Vendor: Silero
What it does: detects speech so transcription runs only when the narrator is talking, which raises the stream count a server can carry substantially.
| 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 (camera streams handled at 25 frames per second) | ≈ 60 streams | ≈ 200 streams | ≈ 600 streams |
The right chain depends on what shares the soundtrack with the narrator and how quickly the text is needed. Our consultants measure word error rate on your own feeds, then recommend the separation and transcription models and the delay they imply.
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 Narrator extraction - transcription. 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.