Key takeaways
- Wikipedia lists what makes speech recognition harder: accent, pronunciation, volume, speed, background noise, echoes and recording characteristics.
- Accuracy is usually measured by word error rate: substitutions, deletions and insertions divided by the number of words actually spoken.
- Spontaneous speech is harder than read speech, and large or confusable vocabularies are harder than small ones.
- Several factors can be fixed before transcription: a close microphone, a quiet room and one speaker at a time.
- Check names, numbers and technical terms by ear, because they are where errors matter most.
Speech-to-text software can get the bulk of a recording right and still miss the words that matter. Several of the causes Wikipedia lists are things you can control before the recording ever reaches the software. This guide explains what makes recognition fail, how accuracy is measured and the recording habits that make the biggest difference.
What makes recognition hard
Wikipedia's article on speech recognition sets out the problem. Voices vary in accent, pronunciation, articulation, roughness, dialect, nasality, pitch, volume and speed, and speech is distorted by background noise, echoes and recording characteristics. It also lists conditions that change accuracy:
| Factor | Easier | Harder |
|---|---|---|
| Vocabulary | Small, distinct words | Large or easily confused vocabulary |
| Speaker | A system built for one speaker | A system meant for any speaker |
| Style | Read speech | Spontaneous speech, with "um"s and false starts |
| Speech | Isolated words | Continuous, naturally spoken sentences |
| Conditions | Quiet room | Environmental noise, echoes, room acoustics |
You cannot change who is speaking, but you can change almost everything in the last row.
How accuracy is measured
Accuracy is usually expressed as word error rate (WER). Wikipedia explains that it compares the recognised text with what was really said, counting three kinds of mistake:
- substitutions, where one word is replaced by another;
- deletions, where a spoken word is missing;
- insertions, where an extra word appears.
WER is the total of those three, divided by the number of words actually spoken. A WER of 0 is a perfect match; Wikipedia notes it can even exceed 1 if a system inserts many extra words.
A worked example. Someone says "Send the report to Ana by Friday", seven words. The software writes "Send a report to Anna Friday". That is two substitutions ("a" for "the", "Anna" for "Ana") and one deletion ("by"): three errors out of seven words, a WER of about 43% for that sentence. Notice which errors matter: "a" for "the" is harmless, but the name and the missing "by" could cause real confusion. A low average WER does not protect the words you care about most.
Recording habits that matter most
- Get the microphone close. Google's help for its Live Transcribe app gives the same advice: hold the device microphone near the person or sound you want to capture. Being closer makes the voice louder relative to room noise and echo.
- Choose a quiet room. Turn off fans, televisions and music; close windows.
- Avoid echo. Soft furnishings absorb sound; bare rooms with hard surfaces reflect it.
- One voice at a time. Two people talking over each other are hard to untangle when you check the transcript against the audio.
- Speak at a steady pace. Wikipedia lists speed among the ways voices vary; clear, unhurried speech gives the software less to guess.
- Spell unusual names and terms aloud once, or note them as you go.
Where to put the phone
The same principle, closer is better, looks different in each setting:
| Setting | Phone placement |
|---|---|
| Voice memo | Held in front of you, speaking towards the microphone, away from wind |
| One-to-one interview | On the table between you, nearer the person whose words matter most |
| Meeting | In the middle of the table, away from laptops, cups and keyboards |
| Lecture | As near the speaker as you are allowed, never at the back of the room |
| Phone or video call | Recorded at the source where possible, rather than from a speaker across the room |
Wherever it goes, put the phone on something soft, such as a folded cloth, so table knocks and vibrations do not reach the microphone.
Before you start a long recording
Make a ten-second test recording and transcribe it. If the test transcript is poor, move the phone closer or change room before you record an hour of audio. It is much easier to fix the setup than the transcript.
After transcription: check where it counts
Even a good transcript needs a targeted check:
- names of people, places and products;
- numbers, dates, prices and quantities;
- technical terms and acronyms;
- anything you will quote.
The listing for our transcription app gives the same advice: accuracy depends on recording quality, background noise and how clearly people speak, so names, numbers and important details should be checked against the original audio, and quotations especially. Verbatim versus clean verbatim covers how polished the final text should be.
Applying it
The habits apply to any recording: lectures need the microphone near the speaker rather than at the back of the room, and interviews benefit from a test recording and one person speaking at a time. Transcribe Audio to Text PRO, one of our video and utility apps, converts recorded speech into transcripts for exactly these uses; its guide explains the workflow.
Good to know
Frequently asked questions
Why does speech recognition make mistakes?
What is word error rate?
How far should the microphone be for transcription?
Why does transcription get names wrong?
Sources
- Speech recognition (opens in a new tab)Wikipedia — accessed
- Word error rate (opens in a new tab)Wikipedia — accessed
- Use Live Transcribe (opens in a new tab)Google Android Accessibility Help — accessed
About the author
Reign Creative LLC — The independent app studio that builds the apps covered on this site and publishes every article here. On Google Play it publishes as Reign Collective Apps.
We use AI tools to help research, draft and edit our articles. Each article lists the sources its factual claims rely on, and details about our apps come from their live Google Play listings. No article on this site has been reviewed by an independent expert unless the reviewer is named on it.
No person at Reign Creative LLC has read this article in full. It was published after automated checks and an AI editorial review, which are described in our editorial policy and are not the same as a person reading every sentence.
Spotted an error? Tell us — how corrections are handled is set out in our editorial policy.