Skip to content
    Uni-Translate.
    ENFR
    Quote

    PERFORMANCE / BENCHMARKS

    ASR & TTS.
    Results made clear.

    Recognising speech. Generating a voice. Compare each stage, with the conditions of our tests.

    01 / ASR

    From speech to text.

    Two measures: transcription time and word errors.

    Transcription speed

    The same English sample, on the same graphics card.

    9 English clips · RTX 4070 Ti

    Median · lower is faster
    1. Uni-TranslateOur engine adapted for streaming
      73.7 ms
    2. NVIDIA Parakeet TDT 0.6B v3Local BF16 reference
      106.9 ms
    3. Qwen3-ASR 0.6BLocal BF16 reference, eager execution
      338.6 ms
    4. Qwen3-ASR 0.6BLocal FP32 reference, eager execution
      296.6 ms

    4.6× versus Qwen BF16 in this test

    We run the models with our engine adapted for streaming. This chart times their complete transcription of audio already supplied.

    Test conditions and sources

    Uni-Translate, Parakeet and Qwen recognize the same English clips. The chart measures how long each system takes to produce the full transcript from audio that is already available.

    Streaming can require more computation than offline execution of the same implementation, particularly when audio segments are processed again. This overhead depends on chunking and caching; it is not quantified by this comparison, which times complete transcription of audio already supplied.

    Sample
    9 English MInDS14 clips, shared by all four configurations
    Measurements
    180 per Uni-Translate and Qwen configuration; 90 for Parakeet, after warm-up
    Setup
    RTX 4070 Ti, one request at a time
    Date
    October 14, 2026

    Compared with the Qwen BF16 reference in this test. Median time includes audio preparation and the local engine response.

    9/9 identical transcripts between Uni-Translate and both Qwen configurations. This check does not include Parakeet.

    Shared panel of nine out of ten audio clips: one clip is excluded because Parakeet BF16 returns an empty transcript in all ten repetitions. The ratio compares Uni-Translate with Qwen BF16 only. The Qwen references run in eager mode; other optimizations may change the results. These local tests measure neither a cloud service nor the delay from microphone to headphones.

    Qwen referenceParakeet reference

    Recognition accuracy

    WER: word error rate. Lower means fewer errors.
    1. Uni-Translate · live recognitionConfirmation: 480 ms · 176 French/English clips
      4.42 %
    2. NVIDIA Parakeet TDT 0.6B v3Final transcript · full audio · 32 French clips
      5.67 %
    3. Mistral Voxtral RealtimeDelay setting: 480 ms · 32 French clips
      7.80 %
    4. NVIDIA Nemotron 3.5 ASRStreaming · 560 ms lookahead · 32 French clips
      9.34 %

    Different corpora: these scores do not support a direct ranking.

    Test conditions and sources

    October 15, 2026 · 4,409 reference words · 176 tested clips.

    “Four” and “4” refer to the same number. We apply the same normalization to the reference and recognized text when calculating our score. WER is calculated on words actually confirmed and emitted live: 4.07 % in French and 4.99 % in English.

    480 ms is the confirmation setting used in this test. Actual delay varies with the speech and how it is segmented; it does not measure the path from microphone to headphones.

    Uni-Translate: 195 errors across 4,409 words, with numbers normalized. 176 French and English clips, October 15, 2026. Confirmation set to 480 ms; actual delay varies.

    NVIDIA Parakeet TDT 0.6B v3: 48 errors across 846 reference words. 32 French clips, August 2026. Corpus and normalization differ from the Uni-Translate test.

    Mistral Voxtral Realtime: 66 errors across 846 reference words. 32 French clips, August 2026. Corpus and normalization differ from the Uni-Translate test.

    NVIDIA Nemotron 3.5 ASR: 79 errors across 846 reference words. 32 French clips, August 2026. Corpus and normalization differ from the Uni-Translate test.

    Test scopes: Uni-Translate, 176 French and English clips with normalized numbers; Parakeet, Mistral and Nemotron, 32 French clips from August 2026, using a different normalization. These are reference points from separate tests, not a like-for-like ranking. WER measures word recognition before translation and voice generation. Parakeet is evaluated on its final transcript from the full audio. Mistral settings with a longer delay are excluded.

    Parakeet referenceMistral referenceNVIDIA reference

    02 / TTS

    Our engine versus Qwen.

    Same RTX 4070 Ti, same texts, same voice reference.

    LOCAL COMPARISON / ONE VOICE

    Faster than real time.

    0.35 RTF
    our engine · median RTFNative Qwen FP32 : 2.01 RTF
    ConfigurationMedian RTFp95 RTF
    Uni-TranslateOur engine · BF160.350.53
    Qwen3-TTS 0.6B · FP32Native API · complete audio2.012.15
    Qwen3-TTS 0.6B · BF16Native API · complete audio2.162.32

    RTF = generation time / audio duration. Below 1: faster than real time. Values are per channel.

    Protocol and measurements
    Sample
    18 shared texts · 8 FR / 10 EN
    Measurements
    180 per configuration · 10 repetitions per text
    Warmup
    20 calls per configuration, excluded from measurements
    Hardware
    RTX 4070 Ti · 12GB · 4 CPU threads
    Dates
    October 14–15, 2026

    Shared panel of 18 out of 20 texts: two texts are excluded from all four profiles to retain only fully completed generations. Same prepared voice reference; models are already loaded.

    Qwen3-TTS 12Hz 0.6B Base: native FP32 and BF16 references, eager execution with TF32 disabled. Our engine is measured in BF16. Percentiles use the 180 retained calls in each profile.

    The native Qwen API returns the entire audio file at once. Our engine streams PCM blocks. RTF covers complete generation in both cases. Each RTF is calculated for an individual generation before computing the median or p95; it is not the ratio of two medians. The p95 is the value below which approximately 95% of measurements fall.

    Outputs and their durations may differ. These measurements do not establish equivalent voice quality or measure the microphone-to-headset path.

    Download measurements (JSON)

    UNI-TRANSLATE / STREAMING OUTPUT

    One voice. Or four at once.

    First audio block received from our engine, with one or four simultaneous voices.

    VoicesMedianp95
    One voice168 ms361 ms
    Four voices233 ms623 ms

    The first PCM block may contain silence: this is not time to the first audible speech.

    1.33 sto complete audio per channel, with four simultaneous voices · median.

    Protocol and measurements

    All four channels are active in parallel. First block and complete generation are measured per channel, not across the whole batch.

    With four voices, complete-generation p95: 2.75 s.

    Outputs and their durations may differ. These measurements do not establish equivalent voice quality or measure the microphone-to-headset path.

    03 / IN A CONFERENCE

    What about a full speech?

    These tests include the translation pipeline. They measure a different scope from standalone ASR or TTS.

    Continuous speeches · 4 languages

    One speech, four languagesFirst audio
    2.52–2.53 s
    A second continuous speechFirst audio
    2.62–3.55 s
    View test results and conditions

    One speech, four languages

    A 2 min 8 s speech in French.
    English, German, Spanish and Italian, in parallel.

    Time to the first translated voice, depending on the language.
    2.52–2.53 s
    Last output received after the speech ended.
    4.84 s

    Translation follows a near real-time pace in this speech.

    A second continuous speech

    A 2 min 35 s speech in English.
    French, German, Spanish and Italian, in parallel.

    Time to the first translated voice, depending on the language.
    2.62–3.55 s
    Maximum final delay after the speech ended.
    5.57 s

    A near real-time pace is also observed in this second speech.

    Time to the first audio and the delay at the end describe two different points. They do not represent a constant delay throughout the conference.

    Short speeches · 1 language

    32 speeches · French to English

    Median
    4.04 s
    95th percentile
    5.02 s
    Test conditions and sources

    In a separate set of 32 French speech sessions translated into English, time to the first voice is longer. Plan the setup around your event format.

    Half of the speech sessions start within this time. 4.04 s

    About 95% of speech sessions start within this time. 5.02 s

    This test uses a single translation channel. These results are not a like-for-like comparison with the four-language speeches above.

    What these numbers measure

    ASR: supplied audio to complete text. TTS: complete-generation RTF, and time to the first audio block in a separate section. This first block may not yet be audible. Conference: speech sent at its natural pace to the first translated audio.

    Internal measurements from August and October 2026, depending on the section. They do not measure the complete microphone-to-headset path, translation fidelity or voice similarity.

    Recognition measures transcribed words; conference tests measure start time and pace. None of these scores measures translation fidelity or voice similarity, or compares quality with a human interpreter.

    What about your conference?

    Languages, rooms and speech formats need to be planned together. Tell us about your event and we can define the right setup.

    Discuss my event