Speech Recognition Services

OpenAI Whisper and Deepgram are both production-grade ASR systems but optimised for different use cases. Whisper is an open-source model that can be self-hosted (data stays on your infrastructure) or called via the OpenAI API. It has near-human accuracy on English (4.4% WER on standard benchmarks), supports 100+ languages, and is the best choice for batch transcription where latency is not a constraint. Deepgram is a managed API service optimised for real-time streaming transcription delivering partial transcripts with <300ms latency, making it the correct choice for live captioning, real-time agent assist, and voice interfaces where users see transcription as they speak. For batch transcription of meeting recordings or call logs: Whisper. For real-time streaming: Deepgram. ClickMasters uses both depending on the latency requirement.

Speaker diarisation is the process of determining "who spoke when" in a multi-speaker audio recording segmenting the transcript by speaker identity. Without diarisation, a meeting transcript is a single stream of text with no attribution: "The deadline is Friday. What about the API integration? We need to finish that first." With diarisation: "Speaker 1 (CEO): The deadline is Friday. Speaker 2 (CTO): What about the API integration? Speaker 1 (CEO): We need to finish that first." Diarisation is implemented with pyannote-audio (a speaker segmentation model) applied before transcription the audio is segmented by speaker, each segment is transcribed, and the transcript is reconstructed with speaker labels. For meeting intelligence, call analytics, and interview transcription, diarisation is essential without it, the transcript has limited business value.

OpenAI Whisper is fully open-source and can be deployed on your own infrastructure either on-premises GPU servers or within your private AWS/GCP/Azure VPC. Audio never leaves your environment. Deployment options: Whisper served via a FastAPI endpoint on an AWS EC2 G5 instance (GPU-accelerated processes a 60-minute meeting in ~2 minutes), or faster-whisper (a CTranslate2-optimised Whisper implementation 4x faster than the original with the same accuracy) for high-throughput batch transcription. For real-time streaming in a private environment, NVIDIA Riva (enterprise-grade on-premises ASR) or a self-hosted Whisper with streaming chunking can replace Deepgram. ClickMasters deploys self-hosted ASR for healthcare, legal, and financial services clients where audio content cannot be sent to external APIs.