How do I transcribe audio to text in Word?

Microsoft Word can turn spoken audio into written text in several ways, and the right method depends on whether you are dictating live, working from an existing recording, or handling long multi-speaker files. The fastest built-in option is Word's Dictate feature, which transcribes your voice in real time as you speak. For pre-recorded audio, Microsoft 365 subscribers get a Transcribe feature that uploads an audio file and returns a full transcript you can insert directly into your document. If you don't have a Microsoft 365 subscription or need to process hours of interviews, podcasts, or meeting recordings with multiple speakers, dedicated AI transcription services generally deliver better accuracy and speaker labeling than Word's native tools. This guide walks through every method step by step, compares their accuracy and cost, and flags the mistakes that waste the most time.

The Direct Answer: Three Ways to Get Audio into Word

Also worth reading: How did OpenAI transcribe over a million hours of audio data? · What equipment do I need to effectively transcribe audio and video recordings? · How can I make the most of the new audio transcribe feature?

There are three practical routes to transcribing audio into a Word document. First, use Word's built-in Dictate tool (Home tab, then Dictate) to speak into your microphone and watch text appear on the page in real time; this works in Word for Microsoft 365 on Windows, Mac, and Word for the web. Second, use Word's Transcribe feature (available through Word for the web at microsoft365.com, and rolling out across desktop versions) to upload an existing .mp3, .wav, .m4a, or .mp4 file up to roughly 300 MB or five hours of audio per month for Microsoft 365 subscribers; Word processes it server-side and gives you a transcript with timestamps and speaker labels you can add to your document with one click. Third, if you lack a subscription or need higher accuracy on difficult audio, run the file through a dedicated AI transcription service such as Deepgram, Whisper-based tools, ElevenLabs' speech-to-text model, or Mistral's Voxtral, then copy or export the resulting text into Word as a .docx or paste it directly.

The choice matters more than most people expect. Live dictation is essentially free and instant but only works for speech happening now — you cannot dictate over a recording of someone else talking. File transcription handles recordings but is gated behind a Microsoft 365 subscription (roughly $6.99–$9.99 per user per month for Personal and Family plans) and has monthly usage limits. Third-party services range from free open-source options like OpenAI's Whisper running locally to paid APIs charging anywhere from about $0.004 to $0.30 per minute depending on the provider and plan. Understanding which route fits your situation saves both money and hours of cleanup editing later.

Method 1: Using Word's Built-in Dictate Feature

Dictate is the simplest option when the words exist only in your head or in a live conversation you're part of. In Word for Microsoft 365, open your document, go to the Home tab, and click the Dictate button (it looks like a small microphone). A red recording indicator appears, and everything you say is converted to text as you speak. You can issue voice commands such as "new line," "new paragraph," "delete that," and basic punctuation commands like "period" or "comma" — though modern versions also auto-punctuate reasonably well without explicit commands. On Windows, press the Windows key + H as a shortcut from almost any text field, including Word.

Accuracy for clear single-speaker dictation in a quiet room typically lands above 90 percent word accuracy on modern systems, which matches what independent benchmarks of commercial speech-to-text engines report for clean audio. Performance degrades noticeably with background noise, strong accents, technical vocabulary, or names — expect to manually correct proper nouns, industry jargon, and numbers regardless of which engine you use. Dictate requires an internet connection because the audio is processed in Microsoft's cloud; there is no fully offline mode in consumer Word. It also requires microphone permission, so check Windows Settings > Privacy > Microphone or macOS System Settings > Privacy & Security > Microphone if the button does nothing when clicked.

A practical workflow many writers use: draft by voice at speaking speed (roughly 130–150 words per minute versus 40–60 for typing), then edit by keyboard. This hybrid approach often triples first-draft output speed. Keep sentences short and pause briefly between them — recognition accuracy drops when you ramble across clause boundaries, because the language model has less context to disambiguate homophones like "their," "there," and "they're."

Method 2: Transcribing Pre-recorded Files with Word's Transcribe Tool

For audio that already exists — an interview recording, a lecture, a voice memo — Word's Transcribe feature is the native path. Sign in to Word for the web with a Microsoft 365 account, create or open a document, then go to Home > Dictate dropdown > Transcribe. You'll see two options: Upload audio to process an existing file, or Start recording to capture audio directly in the browser (useful for recording an in-person conversation on a laptop). Supported formats include .wav, .mp3, .m4a, and .mp4, with a per-file limit around 300 MB and a monthly allowance of up to five hours of uploaded audio for subscribers.

Processing time depends on file length; a one-hour interview typically takes several minutes to complete. When finished, the Transcribe pane shows the transcript split by speaker, with timestamps for each segment. You can rename speakers (Word guesses labels like "Speaker 1" and "Speaker 2"), play back audio synced to any line to verify wording, and then choose how to insert the result: Add all to document inserts the full formatted transcript, Add section adds just the current segment, or you can copy individual lines. Microsoft documents this workflow in its Microsoft 365 for Journalists training materials, aimed at reporters processing interview recordings.

Be aware of real limitations. Speaker diarization (the automatic separation of who said what) works best with two to four clearly separated voices on decent microphones; crosstalk, phone-call quality audio, and overlapping speech cause segments to merge or misattribute. Timestamps are approximate rather than frame-accurate. And the feature is not available in all regions or languages — English is best supported, while other languages have partial availability depending on your tenant settings. If your audio falls outside these comfort zones, third-party tools usually outperform it.

Method 3: Third-Party AI Transcription Services

When Word's native tools fall short, dedicated transcription services fill the gap. The current market splits into three tiers. Cloud APIs and platforms — Deepgram, ElevenLabs' speech-to-text model, Mistral's Voxtral, AssemblyAI, and similar — offer fast, accurate transcription with features like character-level timestamps and speaker diarization; Voxtral in particular markets itself as transcribing "at the speed of sound," meaning faster-than-real-time processing. Open-source local models, chiefly OpenAI's Whisper and its many fine-tuned variants, run on your own hardware; community reports show local Whisper-class models running shockingly fast on Apple Silicon Macs, making offline transcription viable for privacy-sensitive material. Hybrid human-plus-AI services add professional review on top of machine output, which The New York Times' evaluation of transcription services identified as the strongest option when accuracy genuinely matters, such as for legal or published journalistic work.

Benchmark comparisons give useful calibration. AIMultiple's head-to-head of Deepgram versus Whisper found Deepgram's Nova-family models edging out Whisper on speed and cost efficiency for production workloads, while Whisper remains attractive because it is free and runs locally. Published word error rates (WER) for top-tier engines on clean English audio cluster around 5–10 percent, degrading to 15–25 percent or worse on noisy, multi-speaker, or accented audio. ElevenLabs claims an industry-leading WER in internal benchmarks alongside character-level timestamps — useful when you need word-level sync for subtitles. Whatever service you pick, the export path into Word is straightforward: download the transcript as .txt, .docx, or .srt, open it in Word, and apply your own formatting.

Here is the concrete sequence for the most common scenario — turning a recorded interview into a Word document using Microsoft's native tools. First, confirm you have an active Microsoft 365 subscription, because Transcribe is not available in the perpetual-license Office 2021/2024 editions or the free Word online tier. Second, sign in at microsoft365.com or office.com, open Word for the web, and start a blank document. Third, click Home, open the Dictate dropdown, and select Transcribe. Fourth, click Upload audio and select your file (.mp3, .wav, .m4a, or .mp4 under 300 MB), or choose Start recording to capture new audio through your browser. Fifth, wait for processing — a one-hour file takes roughly three to ten minutes. Sixth, review the transcript in the side pane, rename the detected speakers, and correct obvious errors while listening to playback for anything doubtful. Seventh, click Add all to document to insert the timestamped, speaker-labeled transcript into your page, then format headings, remove timestamps if unwanted, and save as .docx.

If you hit the five-hour monthly cap or the accuracy disappoints, pivot to a third-party service without changing your end goal. Export or convert your audio to a supported format, run it through the service, download the transcript as .docx where available, and continue editing in Word exactly as before. Keeping the final editing environment constant means switching transcription engines costs you nothing downstream.

Common Mistakes That Ruin Transcription Accuracy

The single biggest error is feeding the system bad audio. Speech-to-text engines — Microsoft's included — are trained largely on clean, close-mic speech, so recordings made across a room, on a phone in a pocket, or in a reverberant conference space can see word error rates double or triple. Record close to the speaker, minimize background noise, and prefer lossless or high-bitrate formats; a $20 lavalier microphone improves results more than any software upgrade. The second mistake is ignoring speaker separation: if two people talk over each other, no diarization algorithm will reliably untangle it, so ask participants not to interrupt during recorded sessions.

Third, people skip the review pass entirely. Even a 95-percent-accurate transcript contains roughly one error every twenty words — in a 10,000-word interview that's around 500 potential mistakes, and they cluster predictably around names, numbers, acronyms, and domain jargon. Budget roughly one-third of the audio duration for proofreading against playback; a one-hour interview deserves twenty minutes of checking. Fourth, users misjudge punctuation and formatting expectations: raw transcripts arrive as dense paragraph blocks, so plan time to add paragraphs, headings, and speaker labels if the document will be read by humans rather than searched by machines. Finally, some users attempt to transcribe copyrighted third-party content (podcasts, audiobooks) without rights clearance — transcription creates a derivative text, so make sure you have permission for anything you don't own.

Costs, Limits, and When Each Option Makes Sense

Pricing shapes the decision as much as accuracy does. Word's Dictate and Transcribe come bundled with Microsoft 365 Personal ($9.99/month or $99.99/year) and Family ($12.99/month covering up to six people), so if you already subscribe, the marginal cost of native transcription is zero — subject to the roughly five-hours-per-month upload cap. Standalone AI services price per minute of audio: budget API providers charge around half a cent per minute, mid-tier platforms run two to fifteen cents per minute, and human-reviewed services typically cost $1.00–$3.00 per minute with turnaround times ranging from minutes (AI) to days (human). Local Whisper is free forever but demands capable hardware — realistically a machine with 8 GB+ of RAM, and much better performance on Apple Silicon or a CUDA-capable GPU.

Match the tool to the job. Voice-drafting your own writing? Dictate, today, for free within your existing subscription. One-off interview under an hour? Word Transcribe is adequate and convenient. Weekly podcast episodes, focus groups, legal deposition review, or research with dozens of hours of audio? A paid API or platform pays for itself in saved correction time, and the per-minute math is trivial — even thirty cents per minute costs $18 per hour of audio, far below the value of an hour of your editing labor. Privacy-sensitive material (medical, legal, unreleased corporate content)? Prefer local Whisper or verify the vendor's data-retention policy before uploading; cloud services differ widely in whether they retain audio for model training.

Final Recommendations

For most people asking this question, the honest answer is: try Word's native Transcribe first if you already pay for Microsoft 365, because zero additional cost and one-click insertion into your document is hard to beat for short, clear recordings. Escalate to dedicated AI services when you exceed the monthly cap, need reliable speaker labels on multi-person recordings, require precise timestamps for subtitling, or handle confidential audio that shouldn't leave your machine. Whichever engine produces the first draft, treat it as a rough cut — the last mile of transcription quality still comes from a human ear checking the output against the original audio, and no 2026-era system has eliminated that step.