Contact Centre Speech-to-Text
Transcribes calls in English, Bahasa Malaysia and Manglish.
WER (English)
8.9%
WER (Bahasa Malaysia)
10.4%
WER (code-switched)
13.7%
PII redaction recall
0.997
Real-time factor
0.18
About this model
Transcribes contact centre calls in English, Bahasa Malaysia and the code-switched mix customers actually speak, with speaker separation for agent and customer and timestamps per word. Handles banking vocabulary (DuitNow, FD, ASB, credit card instalment plans) and redacts card and MyKad numbers in the transcript.
Intended use
Transcripts for quality assurance, complaint capture, conduct review and search across calls. Transcripts feed the call conduct analyser and complaint classifier. Not for voice biometrics or identification.
Training data lineage
Fine-tuned from the open whisper-large-v3 model on 1,900 hours of consented, PII-redacted contact centre recordings from contact-centre-audio, stored in BigQuery-linked Cloud Storage in the Malaysia region. Fine-tuning and evaluation ran in Vertex AI custom training; the model is served from a Vertex AI endpoint in-region.
Limitations & bias notes
Word error rate roughly doubles on calls with heavy background noise or speakerphone use. Hokkien, Cantonese and Tamil segments are not transcribed and are marked as [other language]. Redaction misses about 1 in 400 spelled-out digits; transcripts stay Confidential.
Ownership and sensitivity
Who approves access
- Owner, Contact Centre, Siti Hajar Ismail
Business-sensitive. Models and data products scoped to named business units.
Entitlement per business unit, approved by the owner; conditions attach.
- Customer PII
- Contains or processes personal data about customers (PDPA 2010).
Try it
Live sandboxcc_call_specimen_0912.wav
fixtureFeedback
Details
- Updated
- 2026-07-16
- Latest version
- 1.3.0
- Licence
- Group Reuse
- Access
- Open to all BUs
- Framework
- Transformers
- Language
- Bilingual
Trained on
contact-centre-audio →