Definition
Streaming transcription (also called real-time transcription or live speech-to-text) is the process of converting audio from a conversation into text while the conversation is happening. The difference from batch transcription is latency: in streaming, text appears on screen seconds after the words were spoken; in batch, transcription is generated after the session ends. For a live co-pilot, streaming is a technical requirement — it cannot detect a crisis signal in real time if audio is processed post-session.
How it's used
Modern streaming transcription systems use Automatic Speech Recognition (ASR) models that process 100-500ms audio chunks and convert them to text with continuous contextual correction. The resulting text typically appears first as a provisional hypothesis (may change) and then as confirmed transcription.
Quality depends on multiple factors: speech rate, accent, background noise, microphone quality, and the amount of technical vocabulary in the domain (clinical terms, coaching jargon). Systems well-tuned for clinical vocabulary in English produce accuracy above 95% in adequate audio conditions.
In bilingual sessions or sessions with code-switching, streaming transcription must detect language in real time to avoid inappropriately mixing vocabulary systems.
When to apply
Streaming transcription is the technical foundation of any live session assistant. Without it, there is no co-pilot. However, streaming transcription is also useful independently: it allows the professional to consult an exact client quote during the session, without relying on memory, or capture exact moments that may be relevant for the clinical note.
Historical origin
The first real-time speech recognition systems were expensive and required specialized hardware. Democratization came with deep learning models and audio processing advances of the 2010s-2020s. The launch of real-time transcription APIs at accessible prices from 2018-2019 opened the door to integrating streaming transcription into third-party applications.
How CauceOS supports it
CauceOS integrates multilingual streaming transcription (neutral global Spanish, English, and automatic language detection) in all sessions. Text appears in the professional's panel with 2-3 second latency. Transcription includes speaker diarization to distinguish the professional from the client in the textual record.
Related terms
- Diarization — identifies who is speaking within the transcript
- Live co-pilot — the system that uses transcription to generate alerts and suggestions
- Live cross-language translation — extension of transcription to bilingual sessions
References
- Radford, A., et al. (2023). Robust speech recognition via large-scale weak supervision. ICML 2023.
- Bahdanau, D., Cho, K., & Bengio, Y. (2015). Neural machine translation by jointly learning to align and translate. ICLR 2015.