We are developing live conference interpretation and have been impressed by Gemini Live Translation’s natural, consistently paced audio. Two discourse-context issues are limiting its suitability for lectures. Please log these as model-quality feedback and consider a session-context feature request.
Test:
7 September 2026; `gemini-3.5-live-translate-preview`; Gemini Developer API; `@google/genai` 2.21.0; English → Spanish (`es`). We streamed a medical conference lecture by an Indian English speaker: first a roughly 60-second middle excerpt, then the full 945.89-second recording from the opening acknowledgments. Input was mono PCM16, 16 kHz, in 100 ms chunks at speaking speed; output was 24 kHz audio with source/target captions enabled
Observed issue
1. **Audience number is not consistently tracked.** The full-talk Spanish begins with plural “les daré”, then uses singular “viste”, “ves” and “lo verás” when referring to what the conference audience saw or will see. For the intended Latin American Spanish delivery, we want plural audience address, with singular used appropriately for an individual. Some English “you” is generic, so we are not asserting that every singular instance is wrong. The issue is unreliable audience tracking and lack of a way to declare the intended setting.
2. **Named addressee becomes the object of seeing.** Around source 08:40–08:53, the speaker addresses Peter about seeing these conditions in Taiwan. The output caption says “debes estar viendo a Peter en Taiwán también”, meaning seeing Peter in Taiwan. The intended meaning is approximately “Peter, seguramente también ves estos casos en Taiwán.” This issue appeared in both the short excerpt and the full-talk output.
Starting from the introduction did not eliminate these issues. The full run used context resumption at source 295.7, 595.1 and 910.1 seconds, with context compression enabled. The inspected medical passage occurred before the second resumption; the short excerpt showed the issue without rotation. These observations are from emitted captions and contextual source review, not a separate transcription of every generated audio word or a statistically controlled evaluation.
## Feature request: context for the live translation session
Please support optional, bounded context supplied at setup and updated during a session, without requiring a conversational-agent workflow:
- Setting and audience: medical lecture addressing a room; target locale and preferred register; default plural audience address.
- Participants and roles: Peter is a chair/addressee; preserve direct address and allow switching from the room to a named individual.
- Domain terminology: talk title, specialist glossary and preferred translations.
- Fidelity preferences: preserve technical specificity, numbers and negation; do not invent completions when source speech is unfinished.
The context should guide interpretation, remain separate from audio to be translated, survive session resumption, and have documented latency/size limits. These could be structured fields rather than unrestricted system instructions. This is a request for runtime context, not necessarily custom model training.
The current translation guide explicitly excludes instructions and documents no glossary/context field. Is there an existing supported mechanism we have missed, or can this be considered for the translation model’s roadmap?
Documentation: https://ai.google.dev/gemini-api/docs/live-api/live-translate