How to read speech recognition and synthesis benchmarks

Read the question before the score

First ask what enters the test and what comes out. ASR turns speech into text; TTS produces audio from text. A recognition measurement therefore does not give translation accuracy, while a synthesis measurement does not describe the complete delay of a conference. The public benchmarks page presents these areas separately. Preserve that distinction when preparing a comparison or purchasing brief, so a component result does not become an unsupported claim about the complete service.

Understand WER, median and p95

WER relates word errors to a reference transcript; its interpretation depends on the dataset and normalisation, among other conditions. It does not directly measure translated meaning. The median divides observations into two halves; p95 indicates a threshold below which roughly 95% of observations fall. A good central result can therefore coexist with longer waits. Read both values and check the number of cases before drawing a broader conclusion about the experience your audience might receive.

Place the local TTS comparison in context

In the public single-voice RTX 4070 Ti trial, Uni-Translate completes audio at a median of 1,004.254 ms, with p95 at 1,938.305 ms. Qwen3-TTS FP32 records 6,251.361 ms and 13,060.148 ms; BF16 records 6,797.511 ms and 14,923.643 ms. The 6.2-times gain concerns the median against FP32 in this trial. Execution environments and generated durations differ; the comparison does not establish equivalent quality. Keep these qualifications alongside the figures whenever you reuse them.

Separate PCM from speech a listener hears

First PCM data arrives at a median of around 168 ms for one voice and 233 ms for four voices in this protocol. That computing milestone is different from the first audible speech in a headset. It also does not describe physical rack delay. The hardware specifications concern an Orin platform; the measurements cited here use an RTX 4070 Ti. Manufacturer TOPS figures cannot convert those results into event latency or validate a listening installation.

Checklist for a useful decision

Connect each number to your question, then plan a representative listening trial. A figure can help identify what to check without being sufficient to accept a complete installation. Keep the protocol link attached to any value you reproduce in an internal brief or comparison document.

• Identify the component and expected result.
• Name hardware, dataset and configuration.
• Read median and p95 together.
• Examine quality and audio duration separately from speed.
• Arrange physical validation before the event.