Live translation benchmark results
Based on current published market benchmarks and our internal testing, Mach recorded the fastest live translation latency in this comparison: an average 0.47 seconds from a completed spoken phrase to readable translated text. The P95 was 1.43 seconds, and every one of the 174 final source phrases received its matching translation.
The result matches the product experience we are building: short, readable translation cards appear continuously while the conversation keeps moving. Users do not need to stop the stream or wait for a long transcript to finish.
Mach Live Translate delivered sub-second average translation responsiveness with 100% final-turn coverage in our internal 15-minute test.
Built for fast, accurate translation
Languages do not all structure sentences in the same way. Translating too early can change the meaning.
Mach translates once a phrase is clear. The result arrives quickly and is easier to trust and read.
Mach vs Gradium, Gemini, GPT, Google Meet, and Meta
Public figures show Meta SeamlessStreaming at around two seconds, Google Meet targeting two to three seconds, Gemini at 2.9 seconds, Gradium at 3.0 seconds, and GPT Realtime Translate at 3.6 seconds. Mach's observed 0.47-second average is the fastest latency figure in this comparison.
The endpoints are different: Mach measures from phrase detection to translated text, while most competitor reports measure speech-to-speech output. That makes Mach's result especially relevant for live captions and readable translation cards, but not a direct speech-to-speech victory claim.
Reliable during a real continuous stream
Speed matters only when the words keep arriving correctly. During the internal stream, Mach accepted all 44,977 audio frames, delivered 174 out of 174 final translations in order, and dropped no audio because of queue pressure.
Separate conversational quality tests across ten languages and both translation directions passed twice. The production language catalog is broader, so we treat those quality tests as strong early evidence rather than a claim that every language pair performs identically.
Benchmark methodology
We replayed BBC World Service audio at real-time speed for 899.540 seconds through Mach's production live translation system, translating English into German. The benchmark measured the time from each completed source phrase to its matching final translated text.
This was a limited internal benchmark: one continuous run, one language direction, and no independent audit or human reference translation for the radio stream. Competitor numbers come from their published reports and use different datasets and endpoints.
- Start-to-ready: 0.564 seconds.
- Average phrase-to-translation response: 0.466 seconds.
- Estimated P50: 0.35 seconds; estimated P95: 1.43 seconds.
- Final coverage: 174 of 174; audio integrity: 44,977 of 44,977 frames.
- Next: public long-form datasets, voice-to-visible timing, repeated runs, and independent human quality review.
Primary sources
These sources describe the reported comparison figures and the public evaluation methodology. Vendor figures remain vendor-reported unless an independent benchmark is explicitly identified.