Context
Descript was evaluating automatic transcription providers to understand which ASR engine offered the strongest combination of speed, accuracy and affordability for its customers.
← Back to case studies
Case study / Descript
Building the data backend behind Descript’s ASR accuracy benchmark.
Descript needed a practical way to manage speech-recognition benchmark data, review Word Error Rate results and support a public comparison of automatic transcription services. MT Software delivered an Airtable-backed workflow and a custom WER reader application to help the team organize, inspect and publish the evaluation data.
System overview
Workflow model
Benchmark data flow from audio samples to WER results
Input
Audio samples and transcripts
Benchmark audio, reference transcripts and ASR-generated transcripts organized for repeatable comparison.
Analysis
WER reader application
A custom workflow for reading Word Error Rate outputs and comparing transcription accuracy across providers.
Output
Publishable benchmark data
Structured results prepared for review, filtering and publication in Descript’s public ASR accuracy article.
System-level workflow representation showing how benchmark evidence, WER analysis and publishable ASR
results
connect.
Overview
From raw transcription tests to a structured benchmark workflow.
The project supported Descript’s comparison of automatic speech-recognition providers by organizing test data, reference transcripts, ASR outputs and WER results into a workflow the team could inspect and publish with confidence.
Challenge
The benchmark required many audio samples, professional reference transcripts, ASR-generated transcripts and WER outputs to be organized in a way that made the results clear, comparable and publishable.
Solution
MT Software built the Airtable backend structure and a custom WER reader application so Descript could review transcription accuracy data and support the final benchmark publication.
Outcome snapshot
A data workflow that supported a published ASR accuracy benchmark.
The public article presented a transparent comparison of transcription services using Word Error Rate, grounded in representative audio samples and professionally verified reference transcripts.
Audio clips evaluated
The benchmark used about 50 clips, each around 3–5 minutes, covering broadcast, calls, meetings and other customer-like audio.
ASR engines compared
The published results compared providers including Google Speech, Temi, Amazon, Speechmatics, Trint, Microsoft and IBM Watson.
Best average WER reported
Google Speech Video achieved the strongest reported average Word Error Rate in Descript’s 2018 benchmark.
Solution design
A backend and analysis workflow designed around benchmark clarity.
The solution focused on making the evaluation data easier to store, inspect and compare, so Descript could move from raw ASR outputs to a credible public accuracy report.
Airtable backend
Structured backend records for audio samples, reference transcripts, ASR providers, test outputs and benchmark metadata.
WER reader app
A custom application layer for reading Word Error Rate outputs and helping the team inspect accuracy results.
Provider comparison flow
Data structures that supported side-by-side comparison across multiple automatic transcription services.
Publication-ready evidence
Organized benchmark data that could support public reporting, filtering and explanation of the ASR accuracy results.
Workflow model
A system-level view of the benchmark data experience.
The visual model keeps the case technical and credible by representing the evaluation workflow without relying on proprietary screenshots or unsupported interface claims.
Benchmark workspace
Audio samples, provider outputs and Word Error Rate results.
Result analysis
Clear movement from ASR output to comparable accuracy results.
WER outputs are framed around audio samples, reference transcripts and provider records so results can be reviewed consistently.
Evidence structure
Reference transcripts, test transcripts, provider metadata and WER values in one benchmark model.
The structure supports a credible case narrative while avoiding fake screenshots or unsupported production interface claims.
Delivery path
A practical delivery sequence for benchmark data infrastructure.
The work is framed around the real operational path: structure the benchmark data, support WER review, organize provider comparisons and prepare the evidence for publication.
-
Phase 01
Benchmark data mapping
Define the relationships between audio samples, reference transcripts, ASR provider outputs and WER results.
-
Phase 02
Airtable backend setup
Structure the backend tables and fields needed to organize benchmark records and keep evaluation data reviewable.
-
Phase 03
WER reader implementation
Build the custom application workflow used to read WER outputs and support transcription accuracy analysis.
-
Phase 04
Results review and publication support
Prepare the organized benchmark data so the Descript team could review, explain and publish the ASR comparison.
Technical direction
Technical structure that supports benchmark accuracy and data clarity.
The technical direction focused on Airtable-backed data organization, WER analysis and benchmark reporting support, using only confirmed project details instead of unsupported production-stack claims.
Airtable backend
WER reader app
ASR benchmark data
Reference transcripts
Provider comparison
Results reporting
Project takeaway
“The value of the build was in turning benchmark data, reference transcripts and WER outputs into a workflow Descript could use to compare ASR providers with clarity.”
