Speech to text: upload the recording, read the transcript

Updated October 29, 2026 · 6 min read

A fifty-minute interview, a lecture, a podcast, a voice note you left yourself. Converting speech into writing is one of the few tasks you can hand over completely: instead of typing along and missing half of it, the text comes out of the recording.

This page describes what transcription does, what it costs, where it breaks — including the limits other vendors leave out of their descriptions.

When transcribing is worth it

• Interviews and research. Find the quote instead of relistening to fifty minutes. Anyone searching for keywords finds them faster in text than in a waveform.
• Lectures and seminars. Recordings become searchable; notes get written afterwards instead of in parallel.
• Podcasts and video. Transcripts feed subtitles, accessible versions and search visibility — search engines index text, not audio.
• Meetings. A raw transcript is the base for action lists; the editorial cut stays human work.
• Voice memos. Capture thoughts on the move, use them as text later.

How it works

1. Upload the audio file

MP3, WAV, M4A, FLAC, OGG or WEBM, up to 25 MB per file. At 128 kbit/s MP3 that is about 25 minutes — split longer recordings into sections first. That sounds like a limitation and is one: three clean sections beat one huge file where nobody knows where it cut off.

2. Transcription

The recording comes back as text with punctuation and paragraphs. Multi-speaker recordings are supported; speaker changes can be told apart in the output.

3. Review and use

The text is readable and copyable in the browser. For a quote, minutes or anything published, a read-through is always worthwhile — proper nouns and jargon are where errors remain.

What it costs

Billing is per second of audio, rounded up and including pauses. One second costs 0.2 credits.

Credits are bought up front: $1 buys 50, packs start at $9. There is no monthly fee and unused credits never expire — you can finish a project in January and pay for the next one in October. The full price list is on the pricing page.

What decides quality

• Background noise beats microphone quality. A phone recording in a quiet room outperforms a condenser microphone next to a running refrigerator. Fans, traffic and room reverb are what recognition trips on.
• One speaker at a time. When two people talk simultaneously, no transcription separates both cleanly — that is physics, not model size.
• Accent and jargon. Mild regional colouring is unproblematic. Heavy dialect, dense jargon and invented terms produce errors you will not see coming. A one-minute test file shows where you stand before you commit hours.
• Numbers, names, acronyms. Years, street names and company names are the errors a proofread finds — not a listen-through finds.
• Music and jingles. Music is not transcribed as text, but it can bury speech. Check the intro of podcasts that open with music.

What transcription does not do

• No summary. The output is what was said, not what it meant. Summarizing stays your job.
• No certified minutes. This is a working transcript. Court or notarial purposes need a different process.
• No speaker identification. Speaker changes are distinguishable, but speakers are not matched to names.
• No rescue of bad recordings. Very quiet, heavily noisy or clipped files come back as fragments, not prose.

Privacy

The audio file is sent to our technology partner for processing only. It is not sold and not used to train models. What is stored, for how long, and your rights are in the privacy policy. For recordings of conversations, the participants' consent is your responsibility, not the tool's.

Questions people ask before paying

Related: text-to-speech generation, voice cloning from a 10-second sample, and the voice and credits FAQ.

From the blog

• what audio work costs per finished minute
• how credits are calculated across cloning and generation
• consent rules when the recording is of other people

Auf Deutsch: Sprachdateien in Text umwandeln — Transkription auf Deutsch.