New: Voice AI Playbook, 6 short videos →
Introducing Transcription Accuracy: See Exactly What Your Voice Agent Heard

Mohamed Salem
Product @ Tuner

Nearly every customer conversation we have comes back to the same question: how well does my agent understand my callers?
Speech-to-text (STT) accuracy changes from one language to the next, and many callers don't stick to one. They switch between English and Spanish mid-call, or drop in a name from another language.
Accents change the picture too. A model can handle one accent well and struggle with the next. Think regional accents across the UK and US, or callers who speak English as a second language.
Then there's where people call from. A car, a busy street or a speakerphone in a noisy office makes every one of these harder.
All of it ends up in the transcript your agent acts on. When STT hears "fourteenth" as "fourth," your agent books the wrong day, and nothing in your logs looks wrong.
Today we're launching transcription accuracy in Tuner. It shows you exactly what your agent heard, word by word, on your own calls.

Two numbers tell you how good your STT is
When you judge a speech-to-text model, two things matter most.
Accuracy: how closely the transcript matches what the caller said. It's 100% minus the word error rate (WER): misheard, missed and added words, divided by the caller's word count.
Latency: how long it takes to return the final transcript once the caller stops talking. Your LLM can't answer until it arrives, so it adds to every reply.
A fast model that mishears books the wrong day quickly, and an accurate one that's slow keeps callers waiting.
Tuner already tracks response latency on every call, STT included. Transcription accuracy adds the other number.
How accuracy is measured
Transcription accuracy compares what your agent heard with what the caller actually said. Three things make that comparison fair.
A reference you can trust: Tuner drafts a transcript of each call. Your team reviews it once and corrects any mistakes. Every replay is then scored against that same verified transcript.
The same audio every time: Tuner plays the same caller audio into your agent, through your real stack. You score what your agent heard, not an offline transcription of the recording.
Only real mistakes count: case, punctuation and filler words like "um" are ignored, and "$50" matches "fifty dollars."

What Good, Fair and Poor mean
Every score also gets a rating, so you can tell at a glance which calls need a closer look.
Good: 97% and up.
Fair: 93% to under 97%.
Poor: below 93%. On a 131-word call, that's 10 or more mistakes.
Even a good score can hide the one word that mattered. That's why every score comes with a list of what went wrong.
Three kinds of mistakes it catches
That list sorts every mistake into one of three kinds.
Misheard: your agent heard a different word. "Fourteenth" became "fourth," and "Ruiz" became "Reece."
Missed: the caller said it, but it never made it into the transcript, like "in the morning."
Added: your agent picked up words the caller never said, like a "go ahead" from someone in the background. That can read as consent the caller never gave.

What you can do with Transcription Accuracy
New STT models are tested on broad, global datasets, and every provider's benchmark tends to favor its own model. Neither tells you how a model handles your callers: the languages they switch between, their accents, the way they speak and the noise behind them.
So test on your own calls. Because the caller audio is identical on every run, you can change one thing and see what it did.
Try a new model: when a new STT model launches, replay your calls through it. See whether it really hears your callers better, and what it does to latency.
Compare models and build your own benchmark: run different models on the same calls, a few times each, and line up their accuracy and latency. The results become a benchmark built on your data, not someone else's.
Tune your settings: adjust custom vocabulary, endpointing or other STT parameters, then check whether misheard and missed words drop.
Catch regressions: providers update their models, and your own setup changes too. Rerun your calls after every update and you'll see if accuracy slipped.
Get started
Transcription accuracy is live.
Start with the call that made you doubt your STT. Verify it, replay it, and see what your agent really heard.
Try transcription accuracy in Tuner →
Your agent can only act on what it heard. Now you can see exactly what that was.
Nearly every customer conversation we have comes back to the same question: how well does my agent understand my callers?
Speech-to-text (STT) accuracy changes from one language to the next, and many callers don't stick to one. They switch between English and Spanish mid-call, or drop in a name from another language.
Accents change the picture too. A model can handle one accent well and struggle with the next. Think regional accents across the UK and US, or callers who speak English as a second language.
Then there's where people call from. A car, a busy street or a speakerphone in a noisy office makes every one of these harder.
All of it ends up in the transcript your agent acts on. When STT hears "fourteenth" as "fourth," your agent books the wrong day, and nothing in your logs looks wrong.
Today we're launching transcription accuracy in Tuner. It shows you exactly what your agent heard, word by word, on your own calls.

Two numbers tell you how good your STT is
When you judge a speech-to-text model, two things matter most.
Accuracy: how closely the transcript matches what the caller said. It's 100% minus the word error rate (WER): misheard, missed and added words, divided by the caller's word count.
Latency: how long it takes to return the final transcript once the caller stops talking. Your LLM can't answer until it arrives, so it adds to every reply.
A fast model that mishears books the wrong day quickly, and an accurate one that's slow keeps callers waiting.
Tuner already tracks response latency on every call, STT included. Transcription accuracy adds the other number.
How accuracy is measured
Transcription accuracy compares what your agent heard with what the caller actually said. Three things make that comparison fair.
A reference you can trust: Tuner drafts a transcript of each call. Your team reviews it once and corrects any mistakes. Every replay is then scored against that same verified transcript.
The same audio every time: Tuner plays the same caller audio into your agent, through your real stack. You score what your agent heard, not an offline transcription of the recording.
Only real mistakes count: case, punctuation and filler words like "um" are ignored, and "$50" matches "fifty dollars."

What Good, Fair and Poor mean
Every score also gets a rating, so you can tell at a glance which calls need a closer look.
Good: 97% and up.
Fair: 93% to under 97%.
Poor: below 93%. On a 131-word call, that's 10 or more mistakes.
Even a good score can hide the one word that mattered. That's why every score comes with a list of what went wrong.
Three kinds of mistakes it catches
That list sorts every mistake into one of three kinds.
Misheard: your agent heard a different word. "Fourteenth" became "fourth," and "Ruiz" became "Reece."
Missed: the caller said it, but it never made it into the transcript, like "in the morning."
Added: your agent picked up words the caller never said, like a "go ahead" from someone in the background. That can read as consent the caller never gave.

What you can do with Transcription Accuracy
New STT models are tested on broad, global datasets, and every provider's benchmark tends to favor its own model. Neither tells you how a model handles your callers: the languages they switch between, their accents, the way they speak and the noise behind them.
So test on your own calls. Because the caller audio is identical on every run, you can change one thing and see what it did.
Try a new model: when a new STT model launches, replay your calls through it. See whether it really hears your callers better, and what it does to latency.
Compare models and build your own benchmark: run different models on the same calls, a few times each, and line up their accuracy and latency. The results become a benchmark built on your data, not someone else's.
Tune your settings: adjust custom vocabulary, endpointing or other STT parameters, then check whether misheard and missed words drop.
Catch regressions: providers update their models, and your own setup changes too. Rerun your calls after every update and you'll see if accuracy slipped.
Get started
Transcription accuracy is live.
Start with the call that made you doubt your STT. Verify it, replay it, and see what your agent really heard.
Try transcription accuracy in Tuner →
Your agent can only act on what it heard. Now you can see exactly what that was.



