October 2026 Summaries

No summary generated yet.

Deepgram has expanded its Nova-3 Medical speech recognition platform with a multilingual model supporting Dutch, English, French, German, Hindi, Italian, Japanese, Portuguese, Russian, and Spanish, including batch and real-time transcription and automatic switching between supported languages. Designed for medical vocabulary rather than general multilingual speech, the model showed lower average word error rates than Deepgram’s general multilingual model, declining from 6.62% to 4.96% in batch transcription and from 8.32% to 5.86% in streaming tests across more than 115 hours of curated medical audio. It also reduced errors involving clinically important entities such as doses, conditions, procedures, and drugs, with batch entity error rates falling by 37% and streaming rates by 40%. Deepgram reports that the model achieved the lowest average batch word error rate among systems tested across all 10 languages and supported applications including ambient clinical documentation, voice agents, and clinician dictation. The company also upgraded its English-only Nova-3 Medical model, reporting improved transcription and entity accuracy, while directing English-only users to the English setting and multilingual or code-switching use cases to the new multi-language setting.

Warm transfers can reduce the customer frustration and lower satisfaction associated with being transferred by preserving structured context for the receiving human agent, rather than requiring callers to repeat their identity, issue, and prior steps. Effective escalation systems use deterministic rules for requests for a person, repeated failed interactions, out-of-scope matters, fraud, emergencies, and other business-defined risks, while passing verified identity, intent, relevant entities, completed actions, an escalation reason code, and a brief summary. The voice agent should signal escalation through a function call, while server-side orchestration and the telephony platform validate routing, store context, and execute either a direct transfer or a bridged conference handoff. Bridged transfers retain the live media path and allow private briefings, transcription, and smoother transitions, though they cost more and require additional setup. Clear caller narration, hold audio, timeout and callback procedures, consent-aware data redaction, duplicate-transfer prevention, and monitoring whether agents re-ask known information are presented as important measures for ensuring that transfer context reaches the human agent and improves the handoff experience.