Each colour maps a published value to the exact source passage used to support it. Only bounded excerpts are public; administrators can inspect the complete captured source.
Title
Heidi Health fine-tunes NVIDIA Parakeet/Nemotron ASR, cutting clinical transcription latency 75% and costs 64%
…ccessibility How Heidi Cut ASR Costs 64% and Latency 75% with NVIDIA… Home Blog How Heidi Cut ASR Costs 64% and Latency 75% with NVIDIA Nemotron Open ASR Heidi Team July 31, 2026 • 6 min read • Listen • Download PDF Table of Contents…
Description
Heidi Health, which processes 2.4 million clinical consultations weekly across 110 languages, fine-tuned the open NVIDIA Parakeet V2 speech recognition model on 1,500 hours of curated clinical audio using the NVIDIA NeMo framework and 8x H100 GPUs, replacing its closed-source vendor ASR. The in-house Nemotron-based ASR stack cut end-to-end transcription latency from about 3.0 seconds to 0.7 seconds (75%+ improvement), reduced word error rate from 13% to 9.4%, and cut operating costs by 64% by eliminating per-minute API pricing, while giving Heidi full ownership of model weights for data sovereignty.
…with your patients It's like your very own junior resident. Get Heidi free Heidi processes 2.4 million clinical consultations weekly , with our AI Care Partner supporting 200+ medical specialties in 110 languages…
Company
Heidi Health
…with your patients It's like your very own junior resident. Get Heidi free Heidi processes 2.4 million clinical consultations weekly , with our AI Care Partner supporting 200+ medical specialties in 110 languages…
Industry
Healthcare
…processes 2.4 million clinical consultations weekly , with our AI Care Partner supporting 200+ medical specialties in 110 languages. At a scale of over 100 million global sessions thus far - each session represe…
Problem
As Heidi scaled to 2.4 million weekly clinical consultations, relying on closed-source, vendor-based ASR created three critical bottlenecks: monthly transcription costs were projected to quadruple within the year, general models struggled with specialized medical vocabulary causing clinically significant errors, and end-to-end latency of about 3.0 seconds interrupted the physician-patient bond.
…tical bottlenecks with general-purpose ASR providers: Economic sustainability : Monthly transcription costs were projected to quadruple within the year, threatening our ability to make clinical AI accessible. Clinical accuracy gaps : General models often struggled with specialized vocabu…
Solution
Heidi fine-tuned the open NVIDIA Parakeet V2 0.6B TDT model using the NVIDIA NeMo framework on a curated dataset of 1,500 hours of clinical audio with an error-focused pipeline prioritizing rare medical terminology, then trained it in 12 hours across 96 total GPU hours on 8x NVIDIA H100 Tensor Core GPUs using NeMo's streaming data-loading tools.
…t interrupted the physician-patient bond. The Solution: Fine-Tuning Parakeet V2 We selected the open NVIDIA Parakeet V2 0.6B TDT (“Parakeet V2”) model as our foundation. Its Transducer-based architecture (TDT) balances high-fidelity accuracy with th…
Technology
NVIDIA Parakeet V2, NVIDIA NeMo, NVIDIA Nemotron ASR, NVIDIA H100
…dical terminology where base models typically fail. 2. High-throughput training Leveraging 8x NVIDIA H100 Tensor Core GPUs, we fine-tuned the 0.6B parameter model in just 12 hours, completing 56,669 training steps across 96 total GPU hours. With each training…
Headline outcome
…seconds (75%+ improvement) Impact: Sovereign, Strategic, and Financial Control 64% reduction in OpEx : By moving away from per-minute API pricing for English transcription, we decoupled our growth from vendor tax. Full model ownership : Owning our weights allows us to deploy across any region…
Business value
The fine-tuned model won 54.5% of blind evaluations against the prior production baseline, cut word error rate from 13% to 9.4% and improved medical-terms F1 score to 93.5%, reduced end-to-end latency from about 3.0 seconds to about 0.7 seconds (75%+ improvement), and cut operating expenses 64% by moving away from per-minute API pricing while giving Heidi full ownership of model weights.
…on Scale Quality: Outperforming the Baseline In blind side-by-side evaluations, our fine-tuned Parakeet model achieved a 54.5% win rate against our previous production baseline. Metric Heidi fine-tuned Parakeet Base Parakeet v2 Legacy vendor solution Non-cu…
AI capabilities
Speech & Audio AI, AI Model Development & MLOps
…at scale, we needed to move beyond general-purpose ASR models. The solution was fine-tuning NVIDIA Parakeet V2 specifically for medical conversations. This required two critical steps: 1. Precision data curation Using the NVIDIA…
Use case type
Document processing
…at scale, we needed to move beyond general-purpose ASR models. The solution was fine-tuning NVIDIA Parakeet V2 specifically for medical conversations. This required two critical steps: 1. Precision data curation Using the NVIDIA…
