How to Transcribe Audio to Text

To transcribe audio to text, run the recording through a speech recognition tool, then read the draft it returns while you listen to the original. The tool can be free software on your own computer, a cloud service that charges by the hour of audio, or a human transcriptionist. Software produces a draft in minutes. A person is slower and far more expensive, but for a recording that is hard to hear, a human check is the safer route when you need a stated accuracy level. Which route to pick depends on how private the recording is, how clearly it can be heard, and what the transcript is for.

A clear recording can become a useful draft in minutes: transcribe it with AI, then proofread it against the audio

Speech recognition is now accurate enough to produce a useful first draft from a clear recording. The large-v2 model of Whisper, an open speech recognition model from OpenAI, had a word error rate of 4.2 percent on English in the Fleurs benchmark ¹. Word error rate is the share of words a tool gets wrong, so 4.2 percent means about four errors in every hundred words ( word error rate explained). A careful reader can fix that draft quickly, but it is not a finished record.

Accuracy also depends on the language. The same model reached 3.0 percent on Spanish ² and 21.5 percent on Hindi ³, and 39.3 percent on Swahili ⁴. If your recording is in Hindi or Swahili, expect far more correction than the English figure suggests.

The workflow is the same whatever tool you use. The tool writes a draft, and you check that draft against the sound. Proofreading is where the accuracy comes from, so use the draft to save typing, not as a finished record, unless the transcript has low stakes.

Tools differ too. In a 2025 test of one 7-minute-31-second podcast episode, Apple's built-in speech engine had a word error rate of 10.3 percent, and Whisper Large V3 Turbo had 1.5 percent ^{5 6}. The author says the exact numbers would change under slightly different methods ⁷, so read one episode as a warning about variation, not as a ranking of products.

Machine time was short in that test. In the same episode, Apple's engine took 9 seconds and Whisper Large V3 Turbo took 40 seconds ^{8 9}. Those times cover the tool's work only, not your proofreading.

There are three routes: on-device AI, cloud AI services and human transcriptionists

Each route trades privacy, accuracy and cost differently. The table sets out the differences; the numbers come below it.

On-device AI is the free route. Whisper's code and model weights are released under the MIT License ¹⁰, so you can run it on your own computer without paying for the software. The hardware sets the limit. The largest model needs about 10 GB of video memory ¹¹, and the smallest needs about 1 GB ¹². If the recording must not leave your machine, this is the route to take, as explained in the question on offline transcription.

Cloud AI charges by the hour of audio. OpenAI's list price for gpt-4o-mini-transcribe works out to about $0.18 per hour ¹³. AssemblyAI lists its Universal-2 model at $0.15 per hour ¹⁴. Prices differ by model and change often, so check the price page on the day you use the service.

Human services cost far more. Rev lists human transcription at $1.99 per minute, with a 99 percent plus accuracy guarantee and delivery in 12 hours or less ¹⁵. An hour of audio at that rate costs about $119.40, against about $0.27 through OpenAI's gpt-transcribe model, a ratio of 442 times ¹⁶. An accuracy guarantee is a promise in the terms, so read them before you rely on one.

The Whisper authors concluded, from their own tests, that its English recognition is very close to human-level accuracy, though not perfect ¹⁷. Human review still matters when a mistake would be costly.

Built-in Apple tools can transcribe some recordings without extra software

Apple's Voice Memos app on a Mac can recognize and transcribe the speech in an audio recording ¹⁸. Apple describes this feature for macOS Sequoia on a Mac with Apple silicon ¹⁹, and the feature is not available in every country or region ²⁰. If you open a recording made in macOS 14 or earlier, Voice Memos transcribes it automatically if it includes speech ²¹.

To take the text out, select the transcript text you want, Control-click it, then choose Copy, and paste it into another document ²². On an iPad, Voice Memos transcription is available on supported models in English and nine other languages ²³.

Built-in tools are free, which makes them the first thing to try. How to transcribe audio for free covers the free options in more detail. When you need a choice of model, a custom vocabulary or subtitle files, look at Whisper or a paid service instead, as compared in our guide to transcription software.

Yap, which we make, is an on-device dictation app for Mac and Windows, built for dictating into the app you are using rather than for transcribing a finished file. For a free file transcription in the browser, see our audio-to-text page.

Audio quality changes accuracy a lot, even with the same model

The same model gives very different results on different recordings. Whisper Large V2 had a word error rate of 25.5 percent on the CHiME6 test set ²⁴ and 17.6 percent on the CallHome set of conversations between friends and family ²⁵, against 4.2 percent on English in the Fleurs benchmark ¹.

Noise hurts accuracy. A 1996 study found that the performance of a speech recognition system degrades rapidly in the presence of ambient noise ²⁶. Sampling rate matters less, but it still counts. In a preliminary Wall Street Journal task, 16 kHz wideband speech gave a word error rate of 8.12 percent, against 8.67 percent for 8 kHz narrowband speech ^{27 28}. Bauer et al. state that dictation and other large-vocabulary applications principally use at least 16 kHz, especially in noisy environments ²⁹.

How to transcribe low-quality audio goes deeper on the recordings that give the tools the most trouble. Four habits improve the draft before you start:

• Get the microphone close. Apple advises that a headset microphone may help when dictation is used in a noisy environment or a room with excessive echo ³⁰.
• Record at 16 kHz or higher. Whisper resamples all audio to 16,000 Hz ³¹, so a higher recording rate adds nothing for that model, but it avoids losing detail before the file reaches the tool.
• List the names. Whisper's initial prompt can supply custom vocabulary or proper nouns, which makes those words more likely to be transcribed correctly ³².
• Listen to the quiet parts. Whisper can invent text when there is no speech. When non-speech sounds were played to the model, a set of hallucinated phrases appeared frequently ³³. In a 2024 study, roughly 1 percent of Whisper transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ³⁴.

Five steps work with any tool: prepare the file, set the language, transcribe, proofread, export

1. Prepare the file. Export the recording in a format the tool accepts. Which audio formats can be transcribed lists the common ones. OpenAI's transcription API takes seven formats: mp3, mp4, mpeg, mpga, m4a, wav and webm ³⁵. The whisper.cpp command-line example runs only with 16-bit WAV files, so convert other formats first ³⁶. Keep the file within the upload limit, which is 25 MB for OpenAI's API ³⁷.

2. Set the language. In the open-source Whisper command line, the language is left unset by default and the program detects the spoken language itself ³⁸. If you know the language, set it yourself, and check that the tool supports it. Whisper supports 98 languages, and OpenAI says accuracy varies by language ³⁹.

3. Transcribe with a model that fits your hardware. Whisper comes in several sizes, and each one trades speed for accuracy:

The turbo model runs at about 8 times the relative speed of the large model ⁴⁴. Use the large model when accuracy matters more than time and your machine has the memory for it.

4. Proofread against the audio. Play the recording at normal speed, stop at each sentence the draft gets wrong, and correct it. Check names, numbers and dates first, because a single wrong digit can change the meaning. Then check the quiet and silent parts, where a tool may have filled gaps with text that was never spoken.

5. Export and check the timestamps. Whisper's own timestamps are per utterance rather than per word, and can be inaccurate by several seconds ⁴⁵. If you need subtitles, check the timing against the audio before you publish. The guide to timestamps and speaker labels explains the trade-offs, and the SRT file and WebVTT formats hold the timings.

If your computer is the place you want to run the model, the guide to running Whisper locally gives the setup in more detail.

Long recordings are split into chunks, so check the joins and the limits of free tiers

Whisper's models are trained on 30-second chunks of audio and cannot take a longer input as a single piece ⁴⁶. If you split a long file yourself, cut it between sentences. OpenAI advises against splitting in the middle of a sentence, because that can remove context and reduce accuracy ⁴⁷. Check the first few words after each cut against the audio.

Service limits vary by request type. Google's synchronous recognition processes audio of 1 minute or less per request ⁴⁸, while its asynchronous requests take audio of about 480 minutes ⁴⁹. Azure's batch transcription can take up to 24 hours to complete at peak hours ⁵⁰.

Free tiers are trials, and they differ in kind:

Check the limit on the day you start, since these pages change. A lecture is a common case of a long recording, and how to transcribe a lecture covers it step by step. For a one-off recording, a free tier may be enough. For a steady flow of recordings, the per-hour price decides the budget.

Proofreading an AI draft takes a fraction of the time typing it would

A draft from a tool arrives in minutes, while typing the same words by hand takes much longer. A one-hour recording at a speaking rate of 150 words per minute holds about 9,000 words. At the 51.56 words per minute online average for typing, typing those words takes about 175 minutes ⁵⁵, before any checking. The speaking rate is a rough figure, since conversations with pauses and overlapping speakers run at other rates.

How long you spend proofreading depends on the audio. Clear, single-speaker audio needs one careful read against the sound. Noisy or overlapping audio needs replays, and a recording with several speakers needs a check of who said what. The guide to timestamps and speaker labels covers that step. For a short personal memo, a quick read is usually enough. For a formal record, plan for a full listen.

Work through the draft in this order:

• Play the recording at normal speed and correct each sentence that does not match.
• Fix names, numbers and dates, and look them up if you are unsure.
• Check for phrases that repeat or appear in silence, since those are the usual signs of an invented passage.
• Decide whether the transcript needs verbatim detail, such as filler words, or a clean version. The guide to verbatim and clean verbatim explains the choice.

Choose the route by privacy, accuracy and budget, not by brand

The right route follows from three questions: can the recording leave your device, how clear is the audio, and how much will the transcript cost you if it is wrong? The chart compares list prices for the AI models with published per-hour prices.

The chart shows the AI models only. Their prices sit in a narrow band, from $0.15 to $0.36 per hour of audio ^{14 56 57 58 59}. A human service sits far above that band, as shown earlier.

Privacy depends on the provider's settings, not only the route. Google says Cloud Speech-to-Text does not log customer audio or transcripts unless the customer opts into data logging ⁶⁰. OpenAI says data sent to its API is not used to train or improve its models unless the customer explicitly opts in ⁶¹, although its abuse monitoring logs are kept for up to 30 days by default ⁶². Read the data terms for the exact plan you use. On a Mac, Apple says that general text dictation is processed on your device and not sent to Siri servers ⁶³, which is a separate feature from transcribing a file.

Use this table to choose:

Start with the free built-in tool if it covers your need. Move to Whisper when you need control over the model or privacy. Pay for a cloud service when you need volume, and pay for a human when a mistake would cost more than the transcript does.

Where to go next

• Meeting transcription for recorded and live meetings.
• Voice memos and voice notes to text for recordings made on a phone.
• Video to text for recordings that include a picture.
• Transcription software for choosing a tool.
• Human vs AI transcription for the trade-off in more detail.
• Interview transcription for recordings with more than one speaker.
• Word error rate defines the accuracy measure used here.
• Automatic speech recognition explains what the tools do.