September 19, 2026
Introducing the Unified Platform for Testing Speech-to-Text APIs
Test and compare speech-to-text APIs from Google, AWS, Azure, AssemblyAI, Deepgram, Yandex and others on the same audio, from one platform. Measure accuracy, speed, cost and features without building a separate integration for every provider.
What's new
- A lot of STT models
- Rich reporting
- Persistence
Speech To Text Tools has been rebuilt.
What started as a simple way to send the same audio through several speech-to-text APIs has grown into a platform for evaluating, comparing and monitoring STT providers on the data that actually matters: your own audio.
Instead of reading benchmarks, opening five playgrounds and maintaining five test integrations, you can run the same files through multiple providers and compare the results side by side.
One audio file. Multiple speech-to-text APIs.
Upload a recording, select the models you want to test, and run them together.
Today, Speech To Text Tools supports models from providers including the top players.
Different models and configurations can be tested without rewriting your application for every API.
This makes it useful both when choosing a speech-to-text provider and when checking whether the provider already in production is still the right one.
An example of a sales call transcribed:
Results Summary
| Provider | CER Raw | CER Norm | WER Norm | WER Strict | Process | RTFx |
|---|---|---|---|---|---|---|
| assemblyai/universal@us:batch | 20.75% | 18.27% | 32.70% | 16.17% | 10.0s | 0.10x |
| aws/transcribe@us-east-1:batch | 11.97% | 11.33% | 17.61% | 12.28% | 17s | 0.17x |
| azure/fast-default@eastus:batch | 10.23% | 9.71% | 16.04% | 9.28% | 15s | 0.15x |
| elevenlabs/scribe-v2@us:batch | 8.96% | 8.50% | 12.26% | 6.59% | 4.4s | 0.04x |
| speechmatics/enhanced@us1:batch | 11.85% | 11.27% | 19.18% | 11.68% | 9.0s | 0.09x |
| openai/gpt-4o-transcribe@global:batch | 8.38% | 7.86% | 14.47% | 7.19% | 13s | 0.13x |
| deepgram/nova_3@global:batch | 15.38% | 12.43% | 29.56% | 9.28% | 3.6s | 0.04x |
💡 Lower is better. CER = Character Error Rate, WER = Word Error Rate. Norm = Normalized (case-insensitive), Strict = content-only.
Reference vs transcript — diff and normalized texts
Diff (Reference → Hypothesis):
Hello, John, how are. How you doing today? You mailed in my company a postcard a few weeks back requesting information on penny stocks that had huge upside potential with very little downside risk. Does that ring a bell? Yeah, I may have sent something. Okay, great. TheWell, reason for the call today, John, is, something just came across my desk, John. It is perhaps the best thing I’'ve seen in the last six months. If you have 60 seconds, I’'d like to share the idea with you. You got a minute? Actually, I’m really very… The nName of the company, Aerotyine International. It is a cutting edge, high- tech firm out of the Midwest, awaiting imminent patent approval on thea next generation of radar detectors that have both huge military and civilian applications now. R. Now, right now, John, the stock trades over- the- counter at 10 cents a share. And by the way, John, our analysts indicate it could go a heck of a lot higher than that. Your profit on a mere $6,000 investment would be upwards of $60,000!. Jesus! Tez, that’'s my mortgage, man. Exactly. You could pay off your mortgage. This stock will pay off my house? John, one thing I can promise you, even in this market, is that I never ask my clients to judge me on my winners. I ask them to judge me on my losers because I have so few. And in the case of AerotynEritime, based on every technical factor out there, John, we are looking at a grand slam home run. Okay, let’'s do it. I’'ll do four4 grand. $4,000? T, that’'d be 40,000 shares, John. Let me lock in that trade right now and get back to you with my secretary with an exact confirmation. Sound good, John? Yeah, sounds good. Great. Hey, John. T, thank you for your vote of confidence. A and welcome to the Investor’s Center. Yeah, thanks a lot, man. Bye-bye.
Red = deletions from reference,Green = insertions in hypothesis
Reference:
Hello, John, how are you doing today? You mailed in my company a postcard a few weeks back requesting information on penny stocks that had huge upside potential with very little downside risk. Does that ring a bell? Yeah, I may have sent something. Okay, great. The reason for the call today, John, is, something just came across my desk, John. It is perhaps the best thing I’ve seen in the last six months. If you have 60 seconds, I’d like to share the idea with you. You got a minute? Actually, I’m really very… The name of the company, Aerotyne International. It is a cutting edge high-tech firm out of the Midwest awaiting imminent patent approval on the next generation of radar detectors that have both huge military and civilian applications now. Right now, John, the stock trades over-the-counter at 10 cents a share. And by the way, John, our analysts indicate it could go a heck of a lot higher than that. Your profit on a mere $6,000 investment would be upwards of $60,000! Jesus! That’s my mortgage, man. Exactly. You could pay off your mortgage. This stock will pay off my house? John, one thing I can promise you, even in this market, is that I never ask my clients to judge me on my winners. I ask them to judge me on my losers because I have so few. And in the case of Aerotyne, based on every technical factor out there, John, we are looking at a grand slam home run. Okay, let’s do it. I’ll do four grand. $4,000? That’d be 40,000 shares, John. Let me lock in that trade right now and get back to you with my secretary with an exact confirmation. Sound good, John? Yeah, sounds good. Great. Hey, John. Thank you for your vote of confidence. And welcome to the Investor’s Center. Yeah, thanks a lot, man. Bye-bye.
Hypothesis:
Hello, John. How you doing today? You mailed in my company a postcard a few weeks back requesting information on penny stocks that had huge upside potential with very little downside risk. Does that ring a bell? Okay, great. Well, reason for the call today, John, is something just came across my desk, John. It is perhaps the best thing I've seen in the last six months. If you have 60 seconds, I'd like to share the idea with you. got a minute? Name of the company, Aerotine International. It is a cutting edge, high tech firm out of the Midwest, awaiting imminent patent approval on a next generation of radar detectors that have both huge military and civilian applications. Now, right now, John, the stock trades over the counter at 10 cents a share. And by the way, John, our analysts indicate it could go a heck of a lot higher than that. Your profit on a mere $6,000 investment would be upwards of $60,000. Jeez, that's my mortgage, man. Exactly. You could pay off your mortgage. John, one thing I can promise you, even in this market, is that I never ask my clients to judge me on my winners. I ask them to judge me on my losers because I have so few. And in the case of Eritime, based on every technical factor out there, John, we are looking at a grand slam home run. Okay, let's do it. I'll do 4 grand. 4,000, that'd be 40,000 shares, John. Let me lock in that trade right now and get back to you with my secretary with an exact confirmation. Sound good, John? Great. Hey, John, thank you for your vote of confidence and welcome to the Investor Center. Yeah, thanks a lot, man. Bye-bye.
Compares every character exactly as typed, including case, punctuation, and spacing. Used by CER Raw.
Keep your tests as projects.
The rebuilt platform is no longer limited to one-off comparisons.
You can organize audio into projects, keep previous runs and build a dataset representing the speech your application actually handles.
That turns an STT evaluation from a spreadsheet created once during vendor selection into something you can repeat.
New model released? Run the dataset again.
Provider changed its model? Compare it against the previous result.
Considering switching from AWS to Deepgram, Google, Azure or AssemblyAI? Test the change before touching production.
From STT comparison to continuous benchmarking.
Speech recognition models don't stand still.
Providers release new models, update existing ones, change features and adjust pricing. The provider you selected a year ago may no longer be the fastest, cheapest or most accurate option for your workload.
Speech To Text Tools is being built around that problem: one independent layer for testing speech-to-text services continuously instead of evaluating every vendor from scratch.
The platform will continue expanding with more providers, models, regions, benchmark data and automated testing.
But you don't need to wait for that.
Take one real recording from your application and see what the current models actually do with it.