Cost-effective multilingual audio transcription at scale with Parakeet-TDT and AWS Batch

This article demonstrates building a cost-effective, scalable audio transcription pipeline using NVIDIA's Parakeet-TDT model with AWS Batch and GPU instances.

• Parakeet-TDT-0.6B-v3 supports 25 European languages with automatic language detection
• Achieves inference speeds of 0.24 seconds per minute of audio
• Event-driven architecture using EventBridge triggers AWS Batch jobs on S3 uploads
• Costs fractions of a cent per hour: $0.00011 on-demand, $0.00005 with Spot Instances
• Buffered streaming inference enables processing long audio on standard g6.xlarge instances
• Local attention mode supports up to 3 hours of audio on 80GB VRAM
• EC2 Spot Instances provide up to 90% cost savings for stateless, idempotent workloads
• CloudFormation template automates infrastructure deployment with GPU monitoring

This solution provides a practical, cost-efficient alternative to managed ASR services for high-volume multilingual transcription workloads.