You asked for audio transcription to save an afternoon of typing. Now there is a searchable text file of everything every speaker said, sitting in a shared folder with no access rules.
Audio transcription is the conversion of spoken words in a recording into written text. A person can do it by listening and typing. Software can do it with speech recognition, in a fraction of the recording's running time. Either way, you end up with a second file holding the same information as the first.
Tool comparisons list word error rate and turnaround time, but a transcript also needs a retention period and an access list. Decide both before the first transcript is created, because once it sits in a shared folder, anyone with the link can search it.
What is audio transcription used for?
Council clerks produce minutes from planning meetings. When a police force prepares a witness interview for prosecutors, it needs a written record first. An insurer with 4,200 claim calls to review will not listen to every one. Journalists check a quote against the tape.
Two styles exist. Verbatim transcription records every word, including filler and false starts. Clean read transcription removes those and tidies the grammar. Interview transcripts prepared for legal proceedings should be verbatim, because a pause or a false start can matter. Meeting minutes can be clean read.
How does automated speech to text work?
Automated systems, called automatic speech recognition (ASR), work in stages. The software cuts the audio into slices of about 25 milliseconds and turns each slice into a numerical description of its sound. A neural network trained on thousands of hours of speech maps those descriptions to likely words. A language model then picks the word sequence that makes most sense, which is how the system chooses between 'their' and 'there'.
Most systems also add speaker labels, called diarization. It separates Speaker 1 from Speaker 2. It does not know their names unless someone supplies them.
Speed is the real gain. Manual transcription typically takes a person four to six hours for each hour of clear audio. Software can return text for a one hour recording in roughly 5 to 15 minutes, depending on the tool and the file size. What you give up is certainty, and the output stays a draft until someone checks it.
Where does audio transcription accuracy break down?
Accuracy is measured as word error rate (WER), the share of words that are wrong, missing or added. To calculate it, add the substitutions, deletions and insertions, then divide by the number of words in a reference transcript that a person has checked. A 5% WER sounds good. On a 30 minute interview at about 130 words a minute, that is roughly 3,900 words and about 195 errors. Some will be harmless. One might drop a 'not' and reverse a statement.
Accents and speech patterns
Models learn from the speech they were trained on. A 2020 Stanford study of five commercial systems, published in PNAS found average error rates of 35% for Black speakers and 19% for white speakers. Regional accents and second language speakers deserve the same test. Run a sample of your own speakers before you trust a tool with a whole archive.
Crosstalk and noise
Two people talking at once is the hardest case. Systems without overlap handling keep one voice and drop the other, or blend both into nonsense. Body-worn camera audio and a meeting room with a ceiling microphone both produce it.
Telephone audio is sampled at 8 kHz, against 16 kHz or more for a decent microphone, so consonants blur. Expect a higher word error rate on phone calls than on microphone audio, so test them separately.
Domain terms and names
Drug names, street names, policy numbers and surnames cause the most trouble. 'Hydroxyzine' can come back as 'hydroxy zine'. Many tools accept a custom vocabulary list, so load one if yours does.
Names matter twice. A misspelled name is still personal data, and a wrong name can attach a statement to the wrong person. No tool reaches 100%. Anything that feeds a decision or a court bundle needs a human read.
Is a transcript a new copy of personal data?
Yes, and it carries the same obligations as the audio. Under Article 4(1) of the GDPR personal data is any information relating to an identified or identifiable person. A name spoken in a recording qualifies. So does the same name in a transcript. The transcript is a second, complete copy of the recording in a different format, and it needs its own rules.
Text behaves differently from audio. Nobody can search a 90 minute recording for a surname. Anyone can search a transcript in a second. Text also pastes into emails, case systems and chat tools without anyone noticing it has moved. A recording in a locked evidence store can be well controlled while its transcript sits in a spreadsheet on a laptop.
Cloud transcription services process the audio on the provider's servers, so it leaves your network. Find out where it goes and who can access it. Whether the provider keeps it to train models is a further question. These are contract questions for your data protection officer. For UK organizations, the ICO's guidance sets out the controller and processor duties that apply.
What sensitive data ends up in transcripts?
A claims call that starts with a policy number can end with a medical history. The table below shows what that kind of recording typically holds.
| Data spoken | Example | Why it matters in text |
|---|---|---|
| Name and address | My name is Priya Shah, 14 Elm Road | Identifies a person directly and is searchable |
| Health details | I was diagnosed in March | Special category data under GDPR Article 9 |
| Card details | Card number and security code read aloud | PCI DSS bans storing the security code after authorization |
| ID numbers | Social Security or National Insurance number | Enables identity fraud |
| Other people | My neighbour told me he was fired | Third parties never agreed to be recorded |
Voice counts too. HIPAA's privacy rules, summarized on the HHS HIPAA hub list voice prints among the identifiers removed for de-identification. The GDPR treats voice data as biometric when it is used to identify someone. A transcript drops the voice and keeps the words. The recording keeps both.
A caller describes their sister's debts. The sister agreed to nothing, yet her name and finances are now in a searchable file.
Storage, sharing and disclosure
Storage
Apply the same retention rule to the transcript as to the audio. If your policy deletes recordings after 90 days and nobody deletes the text, the policy has a hole. Our guide to storing audio files securely covers formats and access controls for the recording itself.
Sharing
Send the least that does the job. A claims handler may need three paragraphs, not a 40 page transcript. Use access-controlled links that expire, and keep transcripts out of email attachments.
Disclosure
Under GDPR Article 15, a person can ask for a copy of the personal data you hold about them. A call recording and its transcript are both in scope, and the voices of other people need a decision before release. Our page on handling subject access requests explains how redaction fits that process.
US agencies face the same pressure through public records requests for interview recordings, under FOIA at federal level and state laws below it. Check your state's exemption for active investigation records before you release anything.
If a transcript goes into evidence, record which audio file it came from, who edited it and when. A chain of custody audit trail keeps that record.
Where do transcription and redaction fit together?
If you redact only the transcript, the name stays in the audio. The text of an audio-only redaction still carries it. Both files need the same treatment.
One workable order is to transcribe first. Then find personal data in the text, remove those time ranges from the audio and regenerate the transcript from the redacted file. Names and numbers are easier to find in text than in a waveform. The timestamps carry each finding back to the audio.
Secure Redact provides audio transcription software and audio redaction software for this workflow. The same recording does not have to move between tools when both steps run in one account. Detection gets weaker as word error rate rises, because a card number with one wrong digit no longer matches a card number pattern. At a 5% error rate per digit, a 16 digit card number has about a 56% chance of containing a wrong digit. Telephone recordings at 8 kHz and lines with heavy background noise are where this shows first. On those recordings automated detection will miss some items, such as a mumbled card number. A person should listen to the redacted file in full before anything is released.
How do you test transcription on your own recordings?
Published error rates describe someone else's audio. Yours will differ, so measure it. Pick ten recordings from your archive and transcribe them. Choose a mix: some phone calls, some room recordings, some body-worn audio if you have it. Correct a five minute stretch of each by hand, count the errors, and divide by the words in that stretch. Compare the results by audio source and by speaker group. A tool that scores well on clean studio audio may score far worse on your phone calls.
Then read the transcripts the way a stranger would: search for a surname and a diagnosis. Whatever turns up is what your retention period and access list have to cover. If the list is longer than you expected, you have your answer on whether redaction belongs in the workflow.
Frequently asked questions
-
Audio transcription is the conversion of spoken words in a recording into written text, done by a person or by speech recognition software.
-
Accuracy depends on the recording. Clean audio with one clear speaker gives the lowest word error rates, while telephone audio, crosstalk, strong accents and specialist terms raise them. At a 5% error rate, a 3,900 word interview contains about 195 mistakes, so any transcript used in a decision or disclosure needs a human check.
-
Yes, if it contains information about an identifiable person, such as a name, address or health detail. Under the GDPR that puts it under the same rules as the recording, including retention limits and subject access requests.
-
Redact both, or redact the audio and regenerate the transcript from the redacted file. Otherwise the information survives in whichever copy you left untouched.