This service takes the sound out of a video file and writes it as an audio file, at the sample rate, channel layout and format the next stage needs — which is the first step of transcription, subtitling, translation, diarization and audio search.
The extraction itself is not a model and needs no GPU: it is a decode-and-write operation, and done correctly it loses nothing. Where models come in is preparation. A recording pulled from video often carries music, effects or room noise that would degrade transcription, and cleaning or separating it first measurably improves everything downstream. Getting the technical details right matters as much: transcription models expect 16 kHz mono, and a track handed over at the wrong rate or with two channels mixed wrongly costs accuracy for no reason.
Models in this group take a recorded video file as input and are applied frame by frame. 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 minute of 1080p video with its audio extracted and prepared.
FFmpeg
Vendor: FFmpeg project
What it does: decodes the video and writes the audio at the required rate and channel layout, without re-encoding where that would lose quality. No GPU needed, and it does the actual extraction.
Requirement
Minimum
Medium
High
GPU type
No GPU required
No GPU required
GPU-accelerated decode (NVENC/NVDEC)
VRAM
—
—
8 GB
vCPUs
2
8
16
RAM
4 GB
16 GB
32 GB
Server
CPU instance, 2 vCPU
CPU instance, 8 vCPU
1× RTX 4090 24 GB
Rate (video minutes processed per hour)
≈ 6,000
≈ 20,000
≈ 60,000
Demucs v4
Vendor: Meta
What it does: separates speech from music and effects, which is what makes a film or broadcast soundtrack transcribable. The best-quality separator we deploy.
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 (video minutes processed per hour)
≈ 360
≈ 1,400
≈ 3,600
MDX-Net
Vendor: Kuielab
What it does: a faster separator with quality close to the best, the economic choice when a whole archive must be processed.
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 (video minutes processed per hour)
≈ 900
≈ 3,300
≈ 8,400
DeepFilterNet 3
Vendor: Friedrich-Alexander-Universität
What it does: removes steady background noise cheaply, which is usually enough for conference recordings and interviews and costs almost nothing.
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 (video minutes processed per hour)
≈ 12,000
≈ 42,000
≈ 108,000
Resemble Enhance
Vendor: Resemble AI
What it does: repairs badly damaged audio from old or low-bitrate video, making a track usable that would otherwise transcribe into nonsense.
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 (video minutes processed per hour)
≈ 480
≈ 1,800
≈ 4,800
Silero VAD
Vendor: Silero
What it does: finds where speech actually occurs, so silent and music-only stretches are dropped before transcription rather than paid for.
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 (video minutes processed per hour)
≈ 36,000
≈ 120,000
≈ 360,000
Choosing between them
What is worth doing depends on your audio and what happens next — cleaning that helps a listener can hurt a transcription model. Our consultants review your files and the downstream task, then recommend the extraction settings and whether a separation or cleaning step earns its cost.
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 Video to audio extraction. 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.