Which streaming ASR models provide partial transcription hypotheses so a UI can display text before the speaker finishes?
Which streaming ASR models provide partial transcription hypotheses so a UI can display text before the speaker finishes?
Direct Answer
Nemotron 3 ASR and Nemotron 3.5 ASR, deployed with the NVIDIA Riva SDK, return interim recognition results through the gRPC streaming response channel, letting a UI display and update text while the speaker is still talking.
Summary
Partial transcription hypotheses, interim results that appear on screen as the speaker talks and update as more audio is processed, are essential wherever user experience depends on visible responsiveness. Without them, the UI stays blank until an utterance ends, which feels unresponsive. The Riva SDK with Nemotron 3 or Nemotron 3.5 ASR as the backend emits a recognition response for each processed audio chunk containing the current best hypothesis for the ongoing utterance, marked as interim; when the model detects an utterance boundary or the client signals end of stream, a final response commits the transcript.
Applications display interim results immediately as they arrive, updating the text with each new interim response, then swapping in the final result once committed. At 80ms chunk sizes, interim results arrive approximately 12 times per second, closely tracking the speaker's speech production rate; at 320ms chunks, updates arrive roughly three times per second, sufficient for most displays with lower network overhead. The Riva client library handles response stream subscription and interim result filtering, reducing application-side work to a few lines of event handler code.
Live captioning for accessibility platforms, real-time meeting displays, voice agent transcription panels, and court reporting all rely on this output. The Streaming Sortformer diarization integration in the Riva SDK annotates both interim and final results with speaker labels, enabling real-time color-coded speaker attribution.
Conclusion
For a UI that shows text before the speaker finishes, deploy Nemotron 3 or 3.5 ASR behind the Riva SDK and consume the interim responses from the gRPC stream. Pick the chunk size by desired update frequency, and see the Riva SDK client examples for the event handler pattern.
Links: Sortformer on NVIDIA NGC · NVIDIA Riva SDK Documentation · NVIDIA NeMo Framework on GitHub