# AI Sales Call Transcription Errors: Why They Happen and What to Ask

Every speech-to-text system mishears some words on a sales call. What matters is whether the tool guesses, and whether you can check the words it relied on.

 By [Russell Pontone](https://quotadot.com/about), Founder & CEO  Published November 16, 2026  7 min read 

Key takeaways

-   Every speech-to-text system mishears some words on a sales call, most often names, product terms and people talking over each other.
-   Word error rate counts inserted, deleted and substituted words, and two vendors can measure it on different audio with different formatting rules.
-   Ask any vendor what its tool does when it is unsure, whether you can see the exact words behind each output, and how you fix a wrong transcript.
-   Quotadot reads whole sentences, stays quiet below a confidence floor, ties each readiness dimension to a verbatim quote, and accepts a corrected transcript for a fresh analysis.

In this article

1.  [Why does speech-to-text mishear sales calls?](#why-does-speech-to-text-mishear-sales-calls)
2.  [How is transcription accuracy measured?](#how-is-transcription-accuracy-measured)
3.  [Why are vendor accuracy claims hard to compare?](#why-are-vendor-accuracy-claims-hard-to-compare)
4.  [What does a misheard word do after the call?](#what-does-a-misheard-word-do-after-the-call)
5.  [What should you ask any vendor?](#what-should-you-ask-any-vendor)
6.  [How does Quotadot handle mishearings?](#how-does-quotadot-handle-mishearings)
7.  [Frequently asked questions](#frequently-asked-questions)

Every speech-to-text system mishears some words on a sales call, most often names, product terms and people talking over each other. No published accuracy number tells you how a tool will do on your own calls. The better question is what the tool does with a mishearing: whether it guesses, and whether you can check the words it relied on.

A transcript feeds everything an AI sales tool produces. The live cue, the summary, the score and the CRM note all start from it. So a buyer is right to ask how good the transcript is, and what happens when it is wrong.

## Why does speech-to-text mishear sales calls?

Microsoft's speech documentation describes call center audio as "Low, 8 kHz, could be two people on one audio channel, could be compressed", with vocabulary that is "Narrow, unique to domain and products" ([Microsoft Learn, December 19, 2025](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-custom-speech-evaluate-data), fetched November 16, 2026). Several of those conditions apply to sales calls. Buyers say company names, product names and people's names. They interrupt. They talk over the rep.

Figure 1

Cause

What it does to the transcript

Source

Names and domain terms

Microsoft lists substitution errors for call center audio as "Medium. Products and people's names can cause these errors"

[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-custom-speech-evaluate-data)

Overlapping speakers, noise

Deletion errors "Can be high. Call centers can be noisy, and overlapping speakers can confuse the model"

[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-custom-speech-evaluate-data)

Weak audio signal

"When you encounter many deletion errors, it's usually because of weak audio signal strength"

[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-custom-speech-evaluate-data)

Who is speaking

Five commercial systems showed "an average word error rate (WER) of 0.35 for black speakers compared with 0.19 for white speakers"

[Koenecke et al., PNAS, 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7149386/)

Words that were never said

In a study of OpenAI's Whisper, "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences"

[Koenecke et al., arXiv, 2024](https://arxiv.org/abs/2402.08021)

Source: each row quotes the linked page, such as [Microsoft's guide](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-custom-speech-evaluate-data), all fetched November 16, 2026. The Microsoft guide is dated December 19, 2025. The Whisper study is version 2 of the arXiv paper, revised May 3, 2024.

Method and limits: the Microsoft rows describe call center audio in general, not sales calls. The Whisper study compared speakers with aphasia to a control group, so its rate describes that audio and that model.

## How is transcription accuracy measured?

The standard measure is word error rate. Microsoft defines it as the count of inserted, deleted and substituted words, divided by the number of words in a human-checked transcript. Its worked example substitutes "Jones" for "John", and it gives that short sentence a word error rate of 60% ([Microsoft Learn, December 19, 2025](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-custom-speech-evaluate-data), fetched November 16, 2026).

The same guide sets rough bands: "A WER of 5-10% is considered good quality and is ready to use. A WER of 20% is acceptable, but you might want to consider more training. A WER of 30% or more signals poor quality." It rates call center audio as "Medium(< 30% WER)".

A low rate can still hide the word that matters. One wrong name in a hundred words is a small error rate and a large error in a CRM note.

## Why are vendor accuracy claims hard to compare?

Two vendors can each publish a rate and still not be measuring the same thing. Speechmatics' accuracy team lead writes that "the problem is harder when comparing across vendors because they all have different ways of formatting" ([Speechmatics, updated September 21, 2025](https://www.speechmatics.com/company/articles-and-news/the-problem-with-word-error-rate-wer), fetched November 16, 2026). He adds that "it is very difficult to get a test set that has no mistakes from human transcribers."

The test audio matters as much as the formatting. A 2024 study of speech recognition services found that "accuracy ranges widely between vendors and for the individual audio samples". It also "measured a significant lower quality for streaming ASR, which is used for live events" ([Kuhn et al., arXiv, August 29, 2024](https://arxiv.org/abs/2408.16287), fetched November 16, 2026). A live coaching tool transcribes as the call happens, so a figure measured on recorded files may not describe it.

Microsoft's own guide asks for "30 minutes to 5 hours of representative audio" to test a model. Your calls are the representative audio.

## What does a misheard word do after the call?

A transcript is the input to every later step. If it puts a competitor's name in the buyer's mouth, a summary built on it can repeat that name, a CRM note can carry it, and a score can count it. A live cue triggered by one keyword can fire on a word the buyer never said.

None of that is visible unless the tool shows you the words it used. A summary with no quotes asks you to trust it. A finding with its quote beside it lets you check it.

## What should you ask any vendor?

1.  What is your word error rate, on what audio, and with what formatting rules?
2.  Was it measured on live streaming audio or on recorded files?
3.  Can we test it on our own calls before we buy?
4.  What does the tool do when it is unsure: guess, or stay quiet?
5.  Can we see the exact words behind each cue, score and note?
6.  How do we fix a wrong transcript, and does the analysis run again?
7.  Which languages does it transcribe?

## How does Quotadot handle mishearings?

Quotadot transcribes on the rep's computer, in English only. The speech-to-text step gives no per-word confidence, so Quotadot cannot mark which single words it was unsure of. It is built to limit the damage a misheard word can do instead.

During the call, the coach labels whole sentences, not single keywords, and it holds a buyer line that is still mid-sentence. When its reading of the buyer falls below a confidence floor, it stays silent. The rep sees nothing rather than a guess. [How Quotadot works](https://quotadot.com/how-it-works)

 and [Live Buyer-State Intelligence](https://quotadot.com/learn/live-buyer-state) explain the live side.

After the call, the [Buyer Readiness Index](https://quotadot.com/learn/buyer-readiness-index) marks each of five dimensions strong or weak only when a verbatim buyer quote supports it, and unproven otherwise. A dimension whose quote is not found word for word in the transcript is dropped. Reps and managers can read the quote and judge it themselves.

If the transcript is badly wrong, a corrected transcript can be pasted on the buyer's page. Quotadot runs the same post-call analysis on it.

The limits, stated plainly:

-   English only.
-   Quotadot does not publish a transcription accuracy benchmark.
-   A live transcript cannot be edited in place. The fix is a corrected transcript pasted afterwards.
-   No audio is stored or sent, only the transcript as text, so there is no recording to replay when you check a word. The [security page](https://quotadot.com/security) covers what is kept.
-   Nothing before the rep presses Start is captured, and a desk phone or mobile call is not heard.

The full answer to this worry sits in the [FAQ](https://quotadot.com/faq#transcription-accuracy).

## Frequently asked questions

### How accurate is Quotadot's transcription, and what happens if it mishears something?

Quotadot uses on-device speech-to-text technology to transcribe English-language sales conversations. Like any AI transcription system, it can mishear words, names, or industry terminology. [Read the full answer](https://quotadot.com/faq#transcription-accuracy).

### What is a good word error rate?

Microsoft's speech documentation calls 5 to 10% good quality and ready to use, 20% acceptable, and 30% or more poor. Those bands describe a whole test set, so test on audio like your own calls.

### Can I compare accuracy numbers from two vendors?

Only with care. Vendors format transcripts differently and test on different audio, and a 2024 study measured lower quality for streaming transcription than for recorded files. Ask each vendor what audio and rules sit behind its number.

### Does Quotadot keep the call audio so I can check a word?

No. Quotadot stores no audio and sends none. It keeps the transcript as text, and each readiness dimension shows the verbatim quote it rests on.

### What if the transcript of a live call is wrong?

A live transcript cannot be edited in place. Paste a corrected transcript on the buyer's page, and Quotadot runs the same post-call analysis on it.

[Russell Pontone](https://quotadot.com/about)

Founder & CEO, Quotadot

Quotadot sells an AI sales copilot and buyer intelligence platform, so the author has a stake in this topic.

Related posts

-   [![](https://quotadot.com/assets/hero-BYO5McRF.jpeg)](https://quotadot.com/blog/real-time-sales-coaching-cognitive-load)
    
    Live coaching
    
    [Real-Time Sales Coaching Without Cognitive Overload](https://quotadot.com/blog/real-time-sales-coaching-cognitive-load)
    
-   [![](https://quotadot.com/assets/hero-Cs3-KF-1.jpeg)](https://quotadot.com/blog/future-of-sales-ai-human-salespeople)
    
    Live coaching
    
    [The Future of Sales AI Should Make Salespeople Stronger, Not Replaceable](https://quotadot.com/blog/future-of-sales-ai-human-salespeople)
    
-   [![](https://quotadot.com/assets/hero-DEYzy1fB.jpeg)](https://quotadot.com/blog/post-call-vs-real-time-sales-coaching)
    
    Live coaching
    
    [Why Post-Call Sales Coaching Is Too Late for the Moments That Decide the Call](https://quotadot.com/blog/post-call-vs-real-time-sales-coaching)
    

[All field notes](https://quotadot.com/blog)

[Quotadot pricing plans](https://quotadot.com/pricing)

Watch one deal run the whole loop.

Fifteen chapters, one buyer, from the first brief to the review after the second call.

[Watch the demo](https://quotadot.com/product-demo)
