
How to Measure Linux Audio Interface Round-Trip Latency
A buffer size is not a round-trip result. Put a signal through a physical output-to-input loopback and record the settings and variation that explain the measurement.
What a loopback result includes
Physical round-trip latency covers the playback and capture paths between the application and the cabled ports. That can include device and driver buffering and, for an analogue loopback, converter delay. A digital loopback answers a different question; do not rank it against an analogue one. There is no verified leaderboard here: the previous figures are not backed by per-run logs and matched repeat measurements available for inspection.
At 48 kHz, a graph set to 128 frames runs a nominal 2.67 ms cycle. Doubling 2.67 ms is not a measurement of the two physical directions: different hardware and drivers may add buffering or have asymmetric paths. A direct-monitor loop is different again, because it can bypass the application. Measure the route a performer needs to hear, not whichever route produces the lowest figure.
Build a fair comparison
- Pick one analogue line output and line input. Start at low level, leave phantom power off on the test input, and disable direct monitoring and other routes that could turn this into a feedback loop. Use the appropriate line-level connections for your device.
- Use a correlation-based loopback measurement client, such as
jack_iodelaywhere packaged. Connect the test client's output to the chosen playback port and the matching capture port to its input. Check that the software sees a clean test signal, not a silent or clipped input. A graph-only self-connection bypasses the hardware and is not a substitute. - Record the observed rate and quantum, including the driver and buffer configuration. Record raw sample results and convert to milliseconds as
1000 × samples / rate. If your tool offers total delay and extra delay, distinguish them rather than mixing the two columns. - Take several runs after the device has settled, and repeat after reopening the stream. Keep the typical result and the spread. Separately note any audible dropouts and the observed error delta; a short low number from an unstable session is not a useful buying result.
Before comparing models, hold the host, connection class, stack, rate, channel count and workload steady where possible. Disclose any difference you cannot control, including hardware revision, firmware and device-specific safety buffers. Save the full tool output, wiring and graph, version manifest, run durations and reset method. A driver or firmware explanation is a hypothesis until you isolate it.
Publish the conditions alongside each result
| Record | Why it matters |
|---|---|
| Interface, revision, firmware and exact ports | A hardware refresh, alternate routing or digital rather than analogue testing can change what is measured. |
| Host, kernel, backend, rate and observed quantum | The same requested buffer can take effect differently on each setup. |
| Test-client version, wiring and alignment settings | A DAW applying recording compensation can make the captured offset look smaller. |
| Raw offsets, repeats, warm-up and workload | A single fast reading tells you nothing about repeatability. |
| Observed audio errors, resets and route changes | A quick result from a broken or restarted graph does not certify stability. |
If each interface works best at a different setting, report both comparisons distinctly: performance at a matched configuration, and performance at each interface's lowest repeatably usable setting for that workload. Never slip one into the other column. If you cannot match the test path, list the results separately and explain why they are not directly comparable. Do not infer timing from a connector version or a published USB speed.
Check capabilities before making a shortlist
Basic recording and playback, independent headphone cueing, mixer/DSP control and firmware servicing deserve separate checks. For example, Focusrite's Linux support guidance distinguishes possible class-compliant operation from official support. Record an unknown as unknown rather than treating a vendor's class-compliance claim as proof every feature works. If you need low-delay dry vocals but not software monitoring, a usable hardware monitor mix may matter more than the cabled round-trip result.
If you are buying, verify supported I/O, mixer access, cueing and firmware needs independently of latency. Likewise, while considering whether to keep a FireWire interface on Linux, compare capability first. A lower delay is not a win if the necessary inputs or monitor mixes no longer work.
- Latency
- USB Audio
- Benchmarks
- Audio Interface
- Linux Audio
- 2026