Transcription Software: How to Choose One

The right transcription software is the one that passes five checks on your own recordings: it processes the audio where your rules allow, it is accurate on your speakers and conditions, it labels speakers and times the text if you need that, it exports the format your next step needs, and its price fits how many hours you transcribe. Brand rankings matter less than these checks, because accuracy differs between vendors and between individual recordings. A 2024 study of eleven speech recognition services on university lecture recordings found that accuracy "ranges widely between vendors and for the individual audio samples" ¹.

This guide walks through the five checks, gives a ten-minute test you can run on any tool, and shows the arithmetic for per-minute, subscription and one-time prices. For the choice between a machine and a person, read Human vs AI transcription. For the whole process from audio to text, read How to transcribe audio to text.

Transcription software comes in four kinds: built-in tools, on-device apps, cloud services and human-backed services

Most transcription software falls into one of four kinds. The kind decides where your audio goes, what it costs and how much you can rely on it for hard recordings.

Built-in tools are the first thing to try. Apple's Voice Memos app on a Mac can recognize and transcribe the speech in a recording ². On macOS Sequoia, Apple describes the feature for a Mac with Apple silicon ³, so check your macOS version and chip before you rely on it. If a tool already on your machine is accurate enough for your audio, you do not need to pay for another.

On-device apps run the model on your own computer, so the audio does not leave the machine. The apps in this group are built on open-source Whisper. Its repository states that the code and model weights use the MIT License ⁴, and the Python package also requires the command-line tool ffmpeg to be installed on your system ⁵. Apps that bundle Whisper may handle that step for you. Buzz describes itself as transcribing and translating audio offline on a personal computer ⁶. Check the memory and speed needs of the model size you pick before you commit. The large model needs about 10 GB of video memory ⁷.

Cloud services are APIs or apps in which the provider runs the model on its servers. AssemblyAI, AWS Transcribe, Google Cloud Speech-to-Text and Azure AI Speech all sell speech-to-text this way, with published prices, and some with free allowances, which the pricing section covers.

Human-backed services send the file to people who check or write the transcript. Rev lists human transcription at $1.99 per minute, with a 99 percent or higher accuracy guarantee and delivery in 12 hours or less ⁸. That guarantee is the vendor's claim, so test it on your own recording like any other tool.

Test accuracy on ten minutes of your own audio before you pay

Run ten minutes of your own audio through each tool you are considering, then score the errors yourself. Published accuracy figures come from other people's recordings, and they vary widely. In the 2024 lecture study, Google's average word error rate on English was 20.1 percent ⁹, while Whisper large-v2 averaged 2.9 percent on the English recordings ¹⁰. Those numbers show the spread. They do not show which tool wins for your audio.

A single independent test shows a similar gap. In a 2025 test by 9to5Mac of one 7-minute-31-second episode, Apple's speech API had a word error rate of 10.3 percent ¹¹, and Whisper Large V3 Turbo had 1.5 percent ¹². One episode is a small sample, so use the result as a reason to test your own audio, not as a verdict.

Accuracy also depends on who is speaking. Across five commercial speech recognition systems, a 2020 study found an average word error rate of 0.35 for Black speakers and 0.19 for white speakers ¹³. Test with speakers who sound like the people in your recordings, including their accents and their rooms. The same study defines word error rate with the total number of words in the correct transcript as its denominator ¹⁴. For more on the measure, see word error rate. Test live mode separately, because a tool that scores well on files can behave differently in a live meeting.

The ten-minute test:

1. Pick ten minutes that include the hardest parts of your work: crosstalk, your accent, a noisy room or a phone line.
2. Make a reference transcript for those ten minutes. Correct a draft by hand, or pay a person to do that segment.
3. Run each tool on the same file with its default settings, and time it.
4. Score the word error rate. Count substituted, deleted and inserted words as errors, and divide by the number of words in the reference transcript.
5. Upload one minute of silence or background noise. If the output contains words, the tool invented them. A 2024 study found that roughly 1 percent of the Whisper transcriptions it tested contained entire phrases or sentences that were not in the audio ¹⁵.
6. Check the speaker labels and timestamps on a two-minute stretch, using the next section as your guide.
7. Multiply the minutes you expect each month by the listed price, using the pricing section below.

Speaker labels, timestamps and an audio-linked editor decide how fast you can correct a transcript

If you need to know who said what, speaker labels matter more than a small gain in accuracy. A transcript with the right words and the wrong speaker tags has to be rebuilt by hand. Google's documentation says its speaker diarization detects when speakers change and labels each voice with a number ¹⁶. Speaker labels can also cost extra. AssemblyAI lists speaker diarization as a separate line at $0.02 per hour of audio ¹⁷, so on a service that bills per hour, the labels add to the transcript rate. Compare the total, not the transcript rate alone. For how labels appear in a finished transcript, see timestamps and speaker labels in transcripts and the glossary entry on speaker diarization.

Timestamps let you find a sentence in the recording again. Check three things in the test:

Where the audio is processed decides who else can access it

Where the audio is processed decides who else can access it. On-device processing keeps the recording on your machine. A cloud service sends it to the provider, and the provider's retention and logging rules then decide how long it stays and who can see it.

On a Mac, Apple says you can check whether general text Dictation is processed on your device and not sent to Siri servers ¹⁸. Build the same habit for any tool: find the statement that says where processing happens, not the summary on the marketing page.

Cloud defaults differ. Google says that by default Cloud Speech-to-Text does not log customer audio or transcripts ¹⁹. Other services may keep transcripts unless you change a setting, so read the retention terms and the setting itself.

For regulated content the bar is higher. In the United States, HHS says that a covered entity and a cloud service provider holding its health information must sign a HIPAA-compliant business associate agreement ²⁰. Confirm that a tool will sign one before you upload patient or client recordings, and check its retention settings in writing.

Before you upload anything sensitive, answer four questions for each tool: Does the audio stay on the device? Is it logged by default? Can you turn retention off? Will the provider sign the agreement your rules require? For dictation rather than file transcription, Yap, which we make, runs on the device on Mac and Windows; see offline dictation on Mac.

Per-minute, subscription and one-time pricing suit different volumes

Per-hour API pricing tends to suit steady bulk work, subscriptions suit regular individual use, and one-time licenses suit people who want to pay once. Free allowances cover occasional use. The table lists the prices this guide checked on 2026-10-17, as published by each vendor or by a US reseller.

The chart puts the four API rates on one scale. The human rate sits far above them: at $1.99 per minute, one hour of audio costs about $119.40 ²⁷. That makes a machine first draft, with a person checking only the segments that must be exact, far cheaper than paying a person for the whole file.

Free allowances differ. AssemblyAI's free tier includes up to 185 hours of pre-recorded transcription ²⁸, which covers a lot of occasional use. Check whether that allowance includes speaker labels, which some vendors price separately.

Subscriptions and one-time licenses need a different calculation. The one-user Dragon Professional v16 license at $685.99 ²⁶ is a one-time price, and the one-year Dragon Professional Anywhere subscription at $660 ²⁵ is a recurring one. On these reseller listings the license costs about one year of the subscription, so it pays off only if you keep using it beyond the first year. Dragon is a dictation product, so read whether Dragon can transcribe audio files before comparing it with file services.

Two rules keep the arithmetic honest. Multiply the hourly rate by the hours you expect each month, then add any extras such as speaker labels. And use the billing you will actually pay, because a plan billed by the year and one billed by the month are not the same price.

Export formats decide what you can do with the transcript afterwards

Export the format your next step needs before you pay. Plain text suits editing, a timed format suits captions and video, and a structured export with word times and speakers suits further processing. The W3C describes a WebVTT file as a sequence of text segments, each associated with a time interval called a cue ²⁹. That timing is why a timed export can feed a video player directly, and a plain text export cannot. For the difference between the formats in practice, see WebVTT, SubRip subtitle files and the guide on captions and subtitles.

Check the input side too. A tool that accepts only a few formats forces you to convert files first. Express Scribe's free version supports common audio formats, including wav, mp3, wma, aif and dct, and its page says the free version does not expire ³⁰. For the everyday case of a voice memo or interview recording, see the guide on voice memos to text.

Manual transcription software still serves people who type transcripts by hand

Manual transcription software plays the audio so you can type it, using keyboard shortcuts or a foot pedal. It does not transcribe anything on its own. It is still the right tool when a person must check every word, and it is also a good way to learn the trade.

Express Scribe is one example. NCH Software says its transcription program uses hot keys or a foot pedal to control audio playback ³¹, and it describes the choice between the two as the ways to run playback ³². A typist can start with the keyboard and add a pedal later. Read the full answer to what Express Scribe is and its alternatives, and the glossary entry on the transcription foot pedal.

Do not confuse transcription with dictation. Dictation is speaking your own text as you write it, and transcription turns a recording you already have into text. The two use related technology, but the test in this guide concerns the second. For the first, see what dictation is. If you are learning the trade of typing transcripts for others, see transcription jobs.

Where to go next

Start with the ten-minute test on the two or three tools you already have, including any built-in tool on your device. Then check the speaker labels, the export format and the processing location for the one you prefer. If you have an interview or lecture to transcribe, the step-by-step guides are the next stop: how to transcribe audio to text, meeting transcription, how to transcribe an interview and transcribing audio with Whisper, ChatGPT or Gemini. Prices change, so re-check the vendor's page on the day you buy. For Yap's own plans, see the pricing page. The transcription overview lists every guide in this series.