Upload a recorded conversation and get readable text back in minutes instead of hours of typing. Picasso IA transcribes interviews, podcasts and calls, splits the speakers apart, and turns an hour of audio into a transcript you can search and quote from.
An hour of recorded conversation is an hour of raw material, not a usable quote. Typing it out by hand takes three or four times as long as the interview itself, and paying a transcription service by the minute adds up fast for anyone doing this every week. Automatic speech to text turns the recording into readable text in minutes, so the work left over is reading, not typing.
Modern speech to text models are trained on enormous amounts of spoken language, and it shows in ordinary conditions: a journalist recording a phone interview, a researcher running a one-on-one session, a podcaster capturing a two-person conversation on decent microphones. In those conditions the model correctly catches most words, follows normal pauses and filler words, and produces a transcript that needs light editing rather than a rewrite. Common languages and mainstream accents, the kind spoken by most guests most of the time, are exactly the case these models were built for.
On Picasso IA, transcription runs from the Toolkit: upload the audio or video file, choose a speech to text model, and the transcript comes back as plain text you can copy, search or edit directly. The same account that transcribes an interview can also generate the images or video for the piece that quotes it, so a single workflow covers both halves of the job.
Automatic transcription is not magic, and pretending otherwise produces embarrassing quotes. Four situations reliably need a second look: unusual names and spellings the model has never seen, technical jargon specific to your guest's field, heavy or regional accents the model was not trained heavily on, and two people talking over each other. None of these make transcription useless, they just mean the first pass is a very good draft, not a finished document.
Pro tip: ask your guest to spell out names, companies and technical terms before the recorded part starts, or leave a ten-second gap where you read them back together. That handful of seconds gives you a reference to check the transcript against, and it is far faster than guessing a spelling from context after the fact.
Overlapping speech is the hardest case for any transcription model, human or automatic. When two people talk at once, the model has to guess which words belong to which speaker, and it sometimes merges both voices into one garbled line. If your interview subject interrupts often, expect to spend more time cleaning those specific passages than the rest of the transcript combined.
A clean transcript is a search index for your own interview. Instead of scrubbing back and forth through an hour of audio trying to remember where someone said the one good line, you search the text for a keyword and land on the exact paragraph. That alone often saves more time than the transcription itself, especially on long recordings where the best quote is buried forty minutes in.
For a written piece, the transcript becomes the raw material for a first draft rather than the draft itself. Read through once for the shape of the conversation, mark the two or three passages that carry the piece, and quote those verbatim rather than paraphrasing from memory. A verbatim quote checked against the transcript is defensible in a way a quote reconstructed from notes is not.
When a recording has more than two people, ask for the transcript split by speaker rather than one continuous block. A labeled transcript, Speaker 1, Speaker 2, Speaker 3, turns a panel discussion or a group interview into something you can actually navigate, instead of a wall of text where you have to guess who is talking by tone alone.
Accuracy depends far more on how the audio was captured than on which model reads it afterward. The same interview recorded in a quiet room and recorded in a busy cafe can produce noticeably different transcripts from the identical model, and knowing that in advance changes where you choose to sit down and record.
| Recording Condition | Expected Accuracy | What Usually Goes Wrong |
|---|---|---|
| Quiet room, single speaker, close mic | Very high | Occasional jargon or unusual names |
| Phone or video call, two speakers | High | Compression artifacts, brief talkover |
| Noisy cafe or open office | Moderate | Words lost under background chatter |
| Overlapping speakers, frequent interruptions | Moderate to low | Merged or dropped lines at crosstalk |
| Strong background music | Low | Lyrics or beat mistaken for speech |
None of these conditions make a recording untranscribable, they just set expectations. A noisy cafe recording is still worth transcribing, the draft will simply need more editing than one recorded in a quiet room.
The process fits comfortably into the same session where you conduct the interview, and most of the time goes into the reading pass rather than waiting on the model.
An honest tool page says where the tool breaks. Automatic transcription is a serious productivity gain, not a replacement for judgment, and five limits show up often enough to plan around.
None of these limits change what the tool is for, turning an hour of audio into readable text quickly; they define where reading carefully still matters. Beyond transcription, the model catalog covers everything from image generation to video, and the Descript comparison shows how a full editing suite compares to a model library built around a single upload-and-generate flow.
For a clear recording of one or two speakers in a normal room, accuracy is high enough that most sentences need no correction at all. What brings accuracy down is not the model's general skill but specific conditions: background noise, overlapping speech, heavy accents and unfamiliar jargon. Treat the first pass as a strong draft and budget a short read-through rather than assuming zero corrections, especially for anything you plan to quote publicly.
Yes, when the recording has reasonably distinct voices and does not overlap constantly. The transcript comes back labeled by speaker turn, which is especially useful for interviews with more than two people, panel discussions or group calls. Overlapping speech is the main thing that confuses speaker separation, since the model has to decide which of two simultaneous voices to attribute a word to.
Mainstream languages and common accents are the strongest case, since that is what the underlying models have seen the most training data for. Regional accents, code-switching between two languages in the same sentence, and less common languages still work but may need more manual correction. If you regularly interview guests in a specific accent or dialect, run a short test recording first so you know what to expect before committing to a full interview.
Processing itself typically takes a few minutes, well under the real-time length of the recording. The time that actually adds up is your own read-through: checking names, fixing a handful of misheard words and splitting speakers if needed. For a clean one-hour interview, budget roughly fifteen to twenty minutes of editing on top of a few minutes of processing.
Yes, and it solves two problems at once. A transcript makes an episode searchable and quotable for show notes or a companion article, and it gives you a text version you can scan for the best clips before editing audio or video. Multi-speaker podcasts benefit the most from speaker-labeled transcripts, since finding who said a specific line becomes a text search instead of scrubbing through audio.
Some background noise is fine and the transcript will still be largely usable, but heavy or constant noise, loud cafes, traffic, music playing under the conversation, measurably lowers accuracy. If you have any control over the recording environment, a quiet room with the microphone close to the speaker will save you far more editing time afterward than any amount of cleanup after the fact.
Not reliably on the first pass, and you should plan for that rather than be surprised by it. Specialized vocabulary, brand names, and names spelled unusually are exactly the words a general-purpose model has seen least often. The fix is simple: keep a short list of the tricky terms and names from your prep notes, and check the transcript against that list before you trust or publish a quote.
Yes, the transcript comes back as plain text, so you can edit it in the Toolkit output, copy it into any writing tool, or paste it into your notes app. There is no locked format or proprietary editor to work around; once the transcription finishes, the text is yours to correct, reformat or trim exactly as you would any other document.
No, and it should not be treated as one. Legal depositions, court proceedings and medical records generally require a certified transcript produced or verified by a qualified human transcriptionist, with a level of accountability an automatic tool does not provide. Automatic transcription is well suited to journalism, research, podcasting and personal notes, where a small correction later carries no legal weight.
New accounts receive free credits, which is enough to transcribe a real interview and judge whether the workflow fits how you work. After that, transcription costs credits like any other generation, and current plans are listed on the pricing page. There is no separate transcription-only tier; the same credits also cover image, video and audio generation across the catalog.
Your next interview is already sitting on your phone as an audio file, and the first transcription costs nothing to try. Upload it to the Toolkit and read the conversation back instead of retyping it.
