Speech transcription and speaker diarisation at call-centre scale

A contact-centre analytics platform serving banks and telcos in Australia and New Zealand needed bespoke training data for an in-house ASR plus speaker diarisation stack – their off-the-shelf cloud transcription was struggling on accented English, code-mixing, and overlapping speech.

Privacy requirements ruled out exporting audio to a generic vendor, and their compliance team had blocked three previous proposals on data-residency grounds.

We provisioned an on-shore-processed annotation pipeline with audio held in the client's Sydney-region storage throughout, accessed by NDA-bound annotators through a hardened VDI. Annotation covered word-level timestamps, speaker labels, sentiment tags, and acoustic events (silence, hold music, IVR transitions).

A weekly calibration session with the client's ML team kept the labelling guidelines aligned with their evolving model behaviour.

Delivered 12,000 hours of transcribed and diarised audio over six months at a measured WER of 4.1% on a 500-hour validation hold-out – beating the client's 5.5% acceptance threshold.

Their production ASR replaced the third-party cloud service three months ahead of plan, cutting per-minute transcription cost by 64% and unblocking a contractual analytics deliverable for a top-three Australian bank.