Speech-to-Text Confidence Scores in Session Transcripts

Published: September 16, 2026

This week's update to the Syllable Agentic Platform captures the confidence the speech-to-text provider reports for each caller turn, stores it with the session, and returns it from the transcripts API. Providers differ in what they report: some give a per-word breakdown alongside the utterance score, others an utterance score alone. Each score therefore travels with the provider and model that produced it, and with whether the provider reported it directly or it was derived, because a confidence value from one provider does not mean the same thing as one from another. When an agent responds in an unexpected way, that detail separates a mis-transcription from a configuration problem, which previously took guesswork to establish. The agent test endpoint also now answers an empty or malformed agent ID with a clear validation error rather than a server error.

Sessions

• Added speech-to-text confidence scores to stored session data and the transcripts API, per utterance and per word where the provider reports a word breakdown, so that an unexpected agent response can be traced to a mis-transcription rather than to configuration.

Agents

• Fixed the Agent Test message endpoint so that an empty or malformed agent ID returns a clear validation error instead of a server error.