---
{
  "title": "How to read speech recognition and synthesis benchmarks",
  "description": "Understand WER, median, p95, first PCM and complete audio measurements in context, without turning a local GPU trial into a performance promise for the rack.",
  "date": "2026-10-24",
  "language": "en",
  "canonical": "https://uni-translate.fr/en/blog/understand-asr-tts-benchmarks/"
}
---

# How to read speech recognition and synthesis benchmarks

A benchmark answers a specific question for a particular dataset, hardware platform and protocol. Separate recognition errors, synthesis speed and end-to-end listening. Public RTX 4070 Ti measurements support a bounded local comparison; they establish neither identical quality nor rack latency. Use their scope to decide what still needs testing for your event.

![A document faces a purely illustrative level relief.](https://uni-translate.fr/ascii-art/svg/measurement-read-report.svg)

## Read the question before the score {#question}
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](https://uni-translate.fr/en/benchmarks/) 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 {#indicateurs}
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 {#comparatif}
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 {#pcm}
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](https://uni-translate.fr/en/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.

![A first-audio measurement does not replace listening for actual audible speech.](https://uni-translate.fr/ascii-art/svg/measurement-ear-and-paper.svg)

## Checklist for a useful decision {#checklist}
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.

## Sources and useful links

- [Rack specifications](https://uni-translate.fr/en/specifications/)
- [Public benchmarks](https://uni-translate.fr/en/benchmarks/)

## Read next

- [Live translation latency: what should you listen for?](https://uni-translate.fr/en/blog/live-translation-latency/)
- [Testing a translation rack: set criteria before the demo](https://uni-translate.fr/en/blog/translation-rack-pilot-criteria/)
- [AI translation or human interpreting: what should you compare?](https://uni-translate.fr/en/blog/ai-translation-human-interpreting/)

By Uni-Translate. Written with AI assistance using public product specifications and documented sources.