About
CES is a Welsh company processing high volumes of customer service calls in both English and Welsh. As call volumes grew, the cost of transcription using AWS Transcribe became unsustainable. Firemind helped CES redesign its speech-to-text architecture using AI operations on AWS, delivering the same accuracy and functionality at a fraction of the cost.
Challenge
CES relied on AWS Transcribe to convert 1,000–2,000 daily calls into text. While accurate, the service became increasingly expensive as call volumes scaled.
Monthly transcription costs ranged between $2,000 and $3,000, driven by AWS Transcribe’s per-minute pricing model.
CES needed to:
- Dramatically reduce transcription costs
- Maintain transcription accuracy and reliability
- Continue supporting both English and Welsh languages
- Preserve advanced capabilities such as speaker separation and PII reduction
- Avoid disrupting existing batch-processing workflows
Firemind was engaged to design and deliver a cost-efficient, production-ready AI operations solution on AWS.
Solution
Firemind replaced AWS Transcribe with a custom, scalable speech-to-text architecture built on AWS managed services:
Deployed an open-source Whisper model (Turbo) on Amazon SageMaker after evaluating alternatives and selecting Whisper for superior Welsh language support
Implemented SageMaker async endpoints with auto-scaling, allowing the platform to scale from zero to active instances based on demand
Optimised infrastructure using ML.G4.XLarge GPU instances with NVIDIA L4 GPUs at approximately $2 per hour
Rebuilt Transcribe-native capabilities using LLMs, including:
Speaker separation
PII detection and reduction
Integrated serverless components, using AWS Lambda for payload processing and S3/SQS for orchestration
Maintained existing workflows, preserving the customer’s four daily batch runs and S3-based ingestion
Delivered full CI/CD automation using AWS CodeBuild for repeatable, production-grade deployments
Results
The new AI operations platform delivered substantial and measurable business impact:
Monthly costs reduced from $2,000–$3,000 to $300–$600
Overall cost savings of 80–85%, equivalent to $1,400–$2,400 per month
Maintained transcription accuracy and feature parity with AWS Transcribe
Continued full support for English and Welsh languages
Scalable, future-ready architecture with potential for an additional 40% cost reduction through further instance optimisation
CES achieved significant cost optimisation without compromising performance, compliance, or language coverage demonstrating how AI operations on AWS can outperform fully managed services at scale.
