Multilingual Transcription at $4 per Hour
A global research and advisory firm transformed its multilingual audio transcription operations with an automated AI pipeline, reducing costs from $600 to $4 per hour of audio and processing time from 24 hours to under one hour.
High Costs and Slow Turnaround Limited Global Reach
The organization relied on external vendors to transcribe specialized meetings, with costs reaching $600 per hour of audio and turnaround times of approximately 24 hours. Limited multilingual support further constrained its ability to serve global clients and expand into new markets.
The client needed a scalable, cost-effective solution capable of processing technical conversations accurately across languages.
Automating Multilingual Transcription With AI
Factored engineered an automated transcription pipeline using Databricks, Apache Spark, and GPU-powered processing.
The architecture combined:
- Speech Recognition: Open-source Whisper and GPT Transcribe models for multilingual transcription.
- LLM Post-Processing: Gemini for transcription refinement, speaker diarization, and technical terminology correction.
- Scalable Orchestration: Spark Streaming, Azure Event Hubs, and Kubernetes for automated processing.
- Cloud Infrastructure: GPU clusters and Azure Blob Storage for efficient audio processing and storage.
The resulting pipeline automated transcription from ingestion through post-processing, reducing reliance on external providers.
Scaling Affordable Transcription Across Languages
The new platform enabled the client to provide fast, affordable transcription across its required languages.
By internalizing transcription capabilities, the organization reduced dependence on costly external vendors while improving its ability to support global clients and expand multilingual services.
Embedding Automated Transcription Into Production
Factored integrated the transcription service into the client's existing workflows, with automated notifications when transcripts were ready.
The rollout included stakeholder validation, cost and accuracy benchmarking, and a phased increase in production volume.
This ensured the solution could operate reliably within the organization's existing processes.
99% Lower Costs and 24× Faster Turnaround
The automated transcription platform delivered measurable improvements:
- 99% Lower Transcription Costs: Reduced costs from $600 to $4 per hour of audio.
- 24× Faster Turnaround: Reduced processing time from 24 hours to under one hour, with P90 processing time of 45 minutes.
- Expanded Multilingual Coverage: Supported transcription across all required languages, enabling broader global service delivery.
These improvements removed operational bottlenecks, reduced vendor dependency, and created opportunities for additional revenue.
Built on Databricks, Spark, and Kubernetes
- Databricks
- Apache Spark
- Kubernetes
- Azure Event Hubs
- Azure Blob Storage
- Whisper
- GPT Transcribe
- Gemini
Engineering Capabilities Behind the Solution
- Data Engineering: Streaming, data collection, integration, and scalable deployment.
- Analytics Engineering: Data modeling and transformation pipelines.
- Machine Learning: Automatic speech recognition, speech-to-text models, and LLM fine-tuning.
- Software Engineering: DevOps and production integration.
- Data Analytics: Stakeholder management and strategic analytics.
