SIGMAAI
Case studies
Case study

Ileum and Ilium Sound the Same to a Model That's Never Heard Either

Domain-vetted transcription and revision for long-form clinical audio, segmented for the conversation instead of the convention.

Ileum and Ilium Sound the Same to a Model That's Never Heard Either
DeliveryPricingCoverageKPIs
Remote, domain-vetted poolPer minute transcribed / reviewedEN-US clinical and radiology audioTranscription throughput, revision throughput, terminology accuracy

The Problem

General-purpose ASR and generalist crowd annotators weren''t built for a radiologist talking through a case out loud.

  • A mistranscribed term isn''t a typo, it''s a different diagnosis. Radiology is full of terms a generic model or an unfamiliar transcriber can mishear as something else entirely, and nothing downstream catches the substitution once it''s written down as fact.
  • Standard segmentation rules assume a structure this audio doesn''t have. Long-form transcription usually caps segments at 30 seconds and applies a 0.5-second gap rule built for multi-speaker, start-stop conversation. A single clinician talking continuously through a single case doesn''t fit that mold.
  • A confident guess reads the same as a correct answer. Nothing in a transcript flags the difference between a term a transcriber knew and one they typed because it sounded plausible.

The Insight

Recognizing that a project needs a different rulebook, and enforcing it before habit takes over, is a human call.

  • The segmentation had to be rebuilt around the audio, not adapted from a template. Each recording is a single speaker working through one continuous case, so the standard 30-second cap and 0.5-second gap rule were dropped for one uninterrupted segment per file, with a 0.1-second buffer applied throughout.
  • Verifying a term is a decision, not a default. Transcribers were required to check uncertain medical language against a clinical reference before submitting, rather than transcribing their best guess and moving on.
  • A transcriber''s first file is a checkpoint, not a formality. Every transcriber''s initial submission went through mandatory review before their work continued unsupervised, catching habits before they became patterns.

The Stakes

  • Experience can work against a project with different rules. Transcribers who already know standard long-form guidelines can default to the 30-second cap or the 0.5-second gap out of habit, producing output that looks correct and fails the project''s actual requirement. Project managers reiterated guidelines and held daily check-ins to ensure quality.
  • A skipped verification hides in plausible-sounding text. A wrong term that sounds right doesn''t announce itself. It reads clean until someone checks it against a reference.

What We Built

  • A segmentation model rebuilt around the conversation. One continuous segment per recording, a 0.1-second buffer applied throughout, and mandatory noise labeling on every file, with the standard long-form segment cap and gap rule explicitly overridden for this project.
  • A two-tier workflow that separated hearing from verifying. A transcription pool converted audio to text; a separate revision pool reviewed that output against guideline and terminology standards before it was considered final.
  • A domain qualification gate before anyone touched the audio. Transcribers had to pass a radiology-specific medical assessment and a language proficiency test, not a general transcription screen.
  • A built-in check against old habits. Every transcriber''s first file, regardless of prior long-form experience, was reviewed against this project''s specific segmentation rules before their work continued unsupervised.
  • A verification step owned by the transcriber, not just the reviewer. Uncertain medical terminology was checked against a clinical reference before submission, with unresolved cases escalated to a Project Manager rather than guessed at.
  • A daily QA cadence sized to a short, high-risk timeline. Project Managers reviewed a sample of the previous day''s completed work every day, catching drift before it compounded across a five-week production window.

Why it Matters Now

  • Speech AI teams training or evaluating clinical language models need audio segmented by the shape of the conversation, not by a template built for interviews or calls. A single continuous clinical exchange loses information the moment it''s artificially split to fit a general-purpose rule.
  • A grounding pipeline is only as reliable as the annotator''s domain fluency. A generalist crowd pool misses the terminology gap a frontier lab can''t catch after the data ships.
  • Segmentation is conversational intelligence made into a repeatable rule. Knowing that one unbroken clinical exchange shouldn''t be cut into arbitrary pieces is the same judgment a person makes instinctively while listening. Here, it''s written into the workflow as a standard, not left to whoever happens to be transcribing that day.

A model doesn''t know it mistranscribed a diagnosis. The person checking it against a reference does.