You have an interview, a lecture, a meeting recording, or a pile of voice memos, and you need it as text — to quote, to search, to caption, or just to skim instead of scrubbing through an hour of audio.
This guide is honest about the options: what manual transcription still gets you, what accuracy AI actually delivers (and where it falls apart), which free tools are already on your devices and where their limits bite, and what the different output formats — plain text, SRT, VTT — are each good for. Then it shows how to transcribe a file in FileMorf, including a trick where the transcript becomes an editor for the audio itself.
Manual vs AI transcription, honestly
For decades transcription meant a human typing while listening, pausing and rewinding constantly. A skilled transcriber takes roughly four times the audio length to produce a clean transcript of an hour — and gets meaning, punctuation, and speaker turns right in ways that still trip up machines. It's slow and, if you pay for it, not cheap.
AI transcription (automatic speech recognition) does it in a fraction of the time for a fraction of the cost. The catch is accuracy, and being clear-eyed about it saves disappointment:
- Clean speech — one clear speaker, good mic, little noise: expect roughly 90–95% accuracy. Very good, but not flawless: plan to proofread.
- Accents and dialects: accuracy drops, sometimes sharply, on accents underrepresented in training data.
- Crosstalk and multiple speakers: when people talk over each other, machines struggle to separate who said what, and words get dropped or merged.
- Background noise: hiss, music, and room echo all pull accuracy down. Noisy audio in means messy text out.
- Jargon, names, and numbers: technical terms, proper nouns, and figures are error-prone exactly because they're unusual — proofread these first.
The realistic workflow
For most people the winning move is AI first, human second: let the machine produce a draft in minutes, then spend a fraction of manual-transcription time correcting it. You get near-manual quality at a fraction of the effort.
Free transcription you already have (and its ceiling)
Before paying for anything, know what's built into devices you own. For live speech, these are genuinely free and good:
- Phone dictation: the microphone key on your iPhone or Android keyboard transcribes as you speak, straight into any text field. Great for a quick note; it's transcribing you, live, into a box.
- Google Docs Voice Typing (Tools → Voice typing) and Microsoft Word Dictate: both transcribe your voice into the document in real time, free, in many languages.
The catch: they don't transcribe files
Every built-in dictation tool listens to a live microphone. None of them will open an existing audio file — an interview you recorded, a downloaded meeting, a voice memo — and transcribe it. The common workaround, playing the file aloud into the mic, degrades quality badly and can't tell speakers apart. To transcribe a file, you need a tool built for files.
Whisper: the open-source watershed
Automatic speech recognition changed sharply when OpenAI released Whisper in 2022. It was trained on a huge, varied set of real-world audio, which made it markedly more robust to accents, background noise, and technical vocabulary than the previous generation — and it was open source, so it spread everywhere.
Whisper is why 'good enough' file transcription became cheap and widely available almost overnight, and it's the engine FileMorf uses for transcription. Because it runs on the whole audio file (not a live mic), it handles the exact case built-in dictation can't: an existing recording, transcribed properly, with punctuation and timing.
Output formats: TXT vs SRT vs VTT
A transcript can come out in several formats, and picking the wrong one means redoing work. The difference is whether the text carries timing:
| Format | Contains | Use it for |
|---|---|---|
| TXT (plain text) | Just the words, no timing | Reading, quoting, searching, pasting into a document, feeding to other software. |
| SRT (SubRip) | Text split into timed cues with start/end timestamps | Subtitles and captions — the most widely supported subtitle format, accepted by YouTube, video editors, and most players. |
| VTT (WebVTT) | Same timed cues, web-native, supports styling/positioning | Captions for web video (the HTML5 <track> element) and modern web players. |
Rule of thumb: if you're going to read or quote it, you want TXT. If it's going on top of a video, you want SRT or VTT. Beyond those, word-level timestamps — a start time for every single word — power features like click-a-word-to-jump and precise transcript-based editing.
Getting better text out: fix the audio first
The single biggest lever on transcript quality is the audio you feed in. No model fully recovers words buried under noise or crosstalk, so a few minutes of prep pays off across the whole transcript:
- Remove noise before transcribing. Hiss, hum, and background chatter all cost you accuracy — clean the audio first, then transcribe the clean version.
- Use the clearest source. If you recorded on multiple devices, transcribe the closest, cleanest mic, not the room recording.
- Trim dead air and cross-talk-heavy sections if you don't need them — less garbage in, less garbage to correct.
- Tell the tool the language. Transcribing English audio as English (rather than auto-detect) avoids the occasional wrong-language guess on short or accented clips.
Noise removal and transcription go together
If a recording is noisy, run voice isolation before you transcribe. Clean audio can be the difference between a 95%-accurate draft you lightly proofread and a 75%-accurate mess you retype.
Transcribing a file in FileMorf
FileMorf's transcription runs on the Whisper engine and is built for files, not a live mic. It produces every format above in one job:
- 1
Upload your recording
Open the transcription tool and add your audio file. Because speech recognition is far too heavy to run in a browser tab, this step runs on secure servers, with the file auto-deleted after the job.
- 2
Get text in every format at once
A single job returns plain text (TXT), subtitle files (SRT and VTT), and word-level timestamp data as JSON. It costs 1 credit per minute of audio; a free FileMorf account includes 6 AI credits per month.
- 3
Proofread the draft
Expect roughly 90–95% accuracy on clean speech. Scan the names, numbers, and jargon first — that's where machine transcription errs — then export the format you need.
The transcript-editing trick
In FileMorf's Audio Studio, the transcript isn't just an output — it's a way to edit the audio. Delete a sentence of text and the corresponding audio is removed too, so you can cut filler words, stumbles, and dead sections by editing text instead of hunting on a waveform. It's the fastest way to tighten a spoken recording.
Transcribe audio to text
Whisper-powered transcription that outputs TXT, SRT, VTT, and word-level timestamps. 1 credit per minute; free accounts include 6 AI credits per month.
Frequently asked questions
Is there a genuinely free way to transcribe an audio file?
Built-in dictation (phone keyboard, Google Docs Voice Typing, Word Dictate) is free but only transcribes a live mic, not existing files. For files, you can run open-source Whisper on your own computer for free, or use a hosted tool with a free tier — FileMorf's free account includes 6 AI credits per month, enough for a few minutes at 1 credit per minute.
How accurate is AI transcription?
Around 90–95% on clean speech from a single clear speaker. Accuracy drops with strong accents, multiple people talking over each other, background noise, and unusual jargon or names. Treat any AI transcript as a strong draft to proofread, not a finished document.
What's the difference between SRT and VTT?
Both are timed subtitle formats and both work for captions. SRT (SubRip) is the most universally supported — YouTube, editors, and most players accept it. VTT (WebVTT) is web-native, supports styling and positioning, and is the format for HTML5 video's <track> element. Use SRT for broad compatibility, VTT for web players.
Can I improve the transcript without retyping it?
Yes — clean the audio before transcribing. Removing noise and crosstalk raises accuracy across the whole file, which means far less to correct afterward. If a recording is noisy, run voice isolation first, then transcribe the cleaned version.
Does transcription upload my audio?
Yes. Speech recognition is too computationally heavy to run in a browser tab, so the file is sent to secure servers for processing and auto-deleted after the job completes. Timeline editing in FileMorf's Audio Studio stays local; only the transcription step uploads.
Keep Reading
How to Remove Background Noise from Audio
Hiss, hum, and background chatter ruin otherwise good recordings. What noise actually is, how to prevent it, the free tools that fix it, and when AI voice isolation wins.
Read the guideHow to Make Slowed and Reverb Edits
The slowed + reverb sound explained: what actually creates it, the two parameters that matter, how to do it in a DAW, and how to make it in one click online.
Read the guide