GoogleAudio modelPreview
Gemini 3.5 Live Translate
PopularityRank 159 of 875
46out of 100same as last month
Summary
Gemini 3.5 Live Translate is a preview model for low-latency speech-to-speech translation. It accepts audio and returns audio rather than treating a general text response as its primary output. Google documents support for more than 70 languages, with output aimed at translated live speech rather than general text reasoning.
The preview endpoint uses the API identifier gemini-3.5-live-translate-preview and returns translated audio plus a text transcript. A separate transcription interface matters only for additional annotations. GPT-Realtime Translate follows a similar speech-to-speech design and also returns transcript deltas.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 19 published benchmark results
Page 1 of 2
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
LiveLingo en-es translation coverage within 30 seconds (Show measurement, test conditions, and source)
- Source value
- 100
- Score
- 100
- Metric
- Share of 90 source clips with a matched translation committed within 30 seconds
- Unit
- %
- Benchmark version
- LiveLingo GPT-Live-1 Interpreter Test, evaluated 2026-09-15; raw results JSON
- Category
- audio-translation-latency
- Direction
- Higher is better
- Sample
- 90
- Source type
- community-reported
- Evaluator
- LiveLingo GPT-Live-1 Interpreter Test
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Community-reported result. LiveLingo GPT-Live-1 Interpreter Test, published and evaluated 2026-09-15. Exact raw model ID gemini-3.5-live-translate-preview. Thirty synthetic English sentences from five domains, joined with 0.5-second gaps and repeated three times per run, yielding 90 clips in a 452,041 ms real-time stream. Identical audio input for all systems; completion lag ends at the last committed translated token relative to source-sentence end. Target pairs en-es and en-ja; output was speech plus transcript.
- Context
- LiveLingo publishes the benchmark and sells its own commercial interpreter, which is included in the comparison; this is not a neutral external replication. The benchmark instance names the exact preview API ID but gives no immutable model revision, complete inference parameters, or confidence intervals. It measures continuous translation coverage and completion latency, not translation quality, recognition accuracy, or speech quality. For en-ja, 89 of 90 translation clips matched; median/p90 lag uses those 89 matches, while coverage uses all 90 clips.
LiveLingo en-ja translation coverage within 30 seconds (Show measurement, test conditions, and source)
- Source value
- 99
- Score
- 99
- Metric
- Share of 90 source clips with a matched translation committed within 30 seconds
- Unit
- %
- Benchmark version
- LiveLingo GPT-Live-1 Interpreter Test, evaluated 2026-09-15; raw results JSON
- Category
- audio-translation-latency
- Direction
- Higher is better
- Sample
- 90
- Source type
- community-reported
- Evaluator
- LiveLingo GPT-Live-1 Interpreter Test
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Community-reported result. LiveLingo GPT-Live-1 Interpreter Test, published and evaluated 2026-09-15. Exact raw model ID gemini-3.5-live-translate-preview. Thirty synthetic English sentences from five domains, joined with 0.5-second gaps and repeated three times per run, yielding 90 clips in a 452,041 ms real-time stream. Identical audio input for all systems; completion lag ends at the last committed translated token relative to source-sentence end. Target pairs en-es and en-ja; output was speech plus transcript.
- Context
- LiveLingo publishes the benchmark and sells its own commercial interpreter, which is included in the comparison; this is not a neutral external replication. The benchmark instance names the exact preview API ID but gives no immutable model revision, complete inference parameters, or confidence intervals. It measures continuous translation coverage and completion latency, not translation quality, recognition accuracy, or speech quality. For en-ja, 89 of 90 translation clips matched; median/p90 lag uses those 89 matches, while coverage uses all 90 clips.
LiveLingo Gemini addendum, recomputed mean comprehension composite per translation case (Show measurement, test conditions, and source)
- Source value
- 4.933333333333334
- Score
- 4.933333333333334
- Metric
- Mean of per-utterance-target two-judge composite ratings (0-5)
- Unit
- rating
- Benchmark version
- LiveLingo Gemini 3.5 Live Translate addendum, evaluated 2026-06-09
- Category
- audio-translation-quality
- Direction
- Higher is better
- Sample
- 120
- Source type
- independent-evaluator
- Evaluator
- LiveLingo raw benchmark results
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Exact raw system ID gemini-3.5-live-translate-preview; 120 utterance-target cases, 30 each to es-ES, zh-CN, ja-JP, and de-DE. The raw rows contain 30 unique source sentence IDs, each evaluated in all four target languages. Recomputed as the arithmetic mean of the 120 per-row means ((GPT-4o judge score + Gemini 2.5 Flash judge score) / 2). The June 9 addendum specifies these two judges; do not apply the main study's separate three-judge description to this addendum. The raw translation field is labeled as a model-provided transcript of its translated speech output, not an independently generated audio transcript.
- Context
- LiveLingo itself sells a competing real-time translation product and publishes this evaluation; its own system appears in the broader comparison. This is a direct commercial conflict, not neutral third-party certification. The Gemini addendum uses two LLM judges (GPT-4o and Gemini 2.5 Flash), while the current main-study method text describes three; these procedures must not be conflated. The judged translation text is the model-provided transcript, so this score does not independently verify that the generated speech matches the text. No immutable provider backend revision or confidence interval is supplied.
LiveLingo Gemini addendum, recomputed median first translated audio (Show measurement, test conditions, and source)
- Source value
- 2.9465
- Score
- 2.9465
- Metric
- Median time from speech onset to first translated audio
- Unit
- seconds
- Benchmark version
- LiveLingo Gemini 3.5 Live Translate addendum, evaluated 2026-06-09
- Category
- audio-latency
- Direction
- Lower is better
- Sample
- 120
- Source type
- independent-evaluator
- Evaluator
- LiveLingo raw benchmark results
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Exact raw system ID gemini-3.5-live-translate-preview. Recomputed median of 120 first_translated_audio_ms values, averaging the two middle values, then converted milliseconds to seconds. This measures first audio from speech onset, not utterance-end-to-output latency.
- Context
- LiveLingo itself sells a competing real-time translation product and publishes this evaluation; its own system appears in the broader comparison. This is a direct commercial conflict, not neutral third-party certification. Speech-onset timing differs from Pinch's TTFA baseline and from utterance-end latency. No immutable provider backend revision or confidence interval is supplied.
Pinch speech translation XCOMET-XL (Show measurement, test conditions, and source)
- Source value
- 0.78
- Score
- 0.78
- Metric
- XCOMET-XL reference-based translation-quality score
- Unit
- No unit provided
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-translation-quality
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. XCOMET-XL is the primary reference-based translation-quality metric.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Pinch speech translation COMET-DA (Show measurement, test conditions, and source)
- Source value
- 0.82
- Score
- 0.82
- Metric
- COMET-DA reference-based adequacy score
- Unit
- No unit provided
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-translation-quality
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. COMET-DA is an auxiliary learned-metric view of translation adequacy.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Pinch speech translation normalized BLEU (Show measurement, test conditions, and source)
- Source value
- 31.31
- Score
- 31.31
- Metric
- Normalized BLEU score
- Unit
- points
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-translation-quality
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. Pinch labels this auxiliary surface-overlap metric normalized BLEU.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Pinch speech translation LongYAAL CU (Show measurement, test conditions, and source)
- Source value
- 1.839
- Score
- 1.839
- Metric
- Computation-unaware average lagging (LongYAAL CU)
- Unit
- seconds
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-latency
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. LongYAAL CU measures sustained lag across the full utterance against an idealized emission schedule and ignores model compute time. Pinch reports 1,839 ms; the stored value converts it to seconds.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Pinch speech translation TTFA p50 (Show measurement, test conditions, and source)
- Source value
- 3.269
- Score
- 3.269
- Metric
- Median wall-clock time from input audio reaching the system to first translated audio chunk (TTFA p50)
- Unit
- seconds
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-latency
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. TTFA p50 starts when input audio reaches the system and ends at the first translated audio chunk. Pinch reports 3,269 ms; the stored value converts it to seconds.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Pinch translated-audio ASR-WER (Show measurement, test conditions, and source)
- Source value
- 0.03
- Score
- 0.03
- Metric
- Word error rate of Whisper transcription against the system's own translated text
- Unit
- ratio
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-intelligibility
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. Pinch transcribes generated speech with whisper-large-v3 and computes WER against the system's own translated text, not a human reference; this measures intelligibility, not translation accuracy. The printed value 0.03 is retained as a ratio, not converted to percent.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Pinch translated-speech UTMOSv2 (Show measurement, test conditions, and source)
- Source value
- 3.64
- Score
- 3.64
- Metric
- UTMOSv2 model-predicted speech naturalness rating
- Unit
- rating
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-naturalness
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. UTMOSv2 is a learned predictor of perceived speech quality applied to translated audio; it is not a human listener panel.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Pinch translated-speech WavLM speaker cosine similarity (Show measurement, test conditions, and source)
- Source value
- 0.7
- Score
- 0.7
- Metric
- Cosine similarity between source and translated-audio speaker embeddings
- Unit
- ratio
- Benchmark version
- Pinch Speech Translation Benchmark, published 2026-06-15
- Category
- audio-speaker-similarity
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Pinch Research
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Independent evaluation. Pinch explicitly names the Gemini 3.5 Live Translate API and evaluates it on kyutai/Audio-NTREX-4L, using identical audio clips and reference translations across French, Spanish, Portuguese, and German into English. The report does not state the tested clip count or immutable API revision. Pinch computes cosine similarity using WavLM fine-tuned for speaker verification and notes it is most meaningful for voice-preserving systems.
- Context
- Pinch is a commercial translation provider and evaluates its own Relay-1 API in the same report. Its authors describe results as directional. This is evaluator-run but commercially conflicted, not neutral third-party certification. The evaluated sample count, provider revision, and uncertainty interval are not reported; do not infer sampleCount from the dataset name.
Page 1 of 2
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
0 to 100, higher is more popular
20 of 83 models
pyannote Community-1, pyannote: 73
Nemotron-3-Diarization, NVIDIA: 57
Gemini 3.5 Transcribe, Google: 52
Confucius4-R2T2, NetEase Youdao: 52
Granite Speech 5.0 470M TurboCTC, IBM: 52
Kokoro-82M, hexgrad: 51
Nova-3, Deepgram: 51
Fish Audio S2 Pro, Fish Audio: 51
NVIDIA Nemotron 3.5 Transcribe, NVIDIA: 50
Granite Speech 5.0 470M TurboCTC NC, IBM: 50
Orukeet, Oruk: 49
Lyria 3.5, Google: 47
Gemini 3.5 Live Translate, Google: 46
Phonon-2, Fermion Research: 44
Aura-2, Deepgram: 44
Gemini 3.1 Flash TTS, Google: 44
MiniMax-Music3, MiniMax: 41
Audio8-ASR-Infinite, Edge0 AI: 41
GPT-Realtime 2.1, OpenAI: 41
Eleven v3, ElevenLabs: 39
Gemini 3.5 Live Translate
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 5, 2026
Gemini 3.5 Live TranslateLyria 3.5Phonon-2Aura-2Gemini 3.1 Flash TTS
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Yes. Google documents SynthID for audio outputs of the listed Gemini Audio models.
- Source: AI watermarking
Licence and access
API
Further details
Sources and data date
Show all 17 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 15, 2026. 9 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | pinch-sample-count Note Pinch does not publish the number of evaluated Audio-NTREX-4L clips for its Gemini results; no sampleCount is assigned to those rows. |
| Unresolved | immutable-provider-revision Note Neither evaluator identifies an immutable Gemini API backend revision for its run. |
| Additional source | Google Gemini API model catalog (retrieved September 22, 2026) · Google Gemini API model catalog · Editorial description reviewed October 2, 2026 |
| Additional source | Official API pricing (retrieved October 25, 2026) |
| Additional source | LiveLingo raw benchmark results (retrieved October 15, 2026) |
| Additional source | Pinch Research (retrieved October 15, 2026) |
| Additional source | LiveLingo GPT-Live-1 Interpreter Test (retrieved October 17, 2026)12 |
| Additional source | Korean Multi-turn Context Translation Benchmark (retrieved October 19, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026)123 |
| Additional source | Google DeepMind Gemini Audio (retrieved November 5, 2026) |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |