How it works
Where your recording goes, and how the results are checked.
From your recording to the page
Your answer is recorded in the browser and sent to Sayfield's server.
Open models write down the words and place each one in time.
Three separate tools listen for pauses; a pause is marked only when two of them agree.
Sayfield's own rules mark the fillers (um, uh and the like) and group the words into phrases. Pitch, how the voice rises and falls, is measured with Praat, a standard tool for studying speech sounds, but isn't shown in the app yet.
What it's built on
First came a review of the research: how unscripted speech is transcribed word for word, how pauses and phrase endings are marked, and what feedback on speaking helps.
One finding shaped the product: speech-to-text tools drop fillers, and in our own test they dropped more of them for people speaking English as a second language. Keeping fillers is the first job for Sayfield's models.
How it's checked
Checked today. Today's words and fillers were checked on public meeting recordings transcribed by hand. The app still misses some fillers, more often for second-language speakers.
Tested next. Sayfield's own models will be tested on recordings they have never seen, against marks checked by people, never against another program's raw output. Results will be reported for each first language, and no number is published until a test backs it.
The notes
After each answer you get two short notes: one line of yours worth keeping, and one thing to try in Take 2 (answering the same question again).
An AI model reads your transcript and Sayfield's measurements of your timing, never your voice. It picks the line of yours to keep, one thing to try and where to try it. It also writes a short sentence about the kept line, an example of the thing to try built mostly from your own words and, now and then, a question to listen with. Every quote is checked word for word against your transcript before you see it. The name of the thing to try, its rule and the line on why it helps are wording Sayfield wrote in advance.
Under each thing to try, one plain line says why it helps. Most of the moves come from research, such as a study of 62 videotaped job interviews (Bangerter and colleagues, 2014), reviews of persuasion studies (O'Keefe, 1997 to 1999), and work on how arguments and spoken stories are built (Toulmin, 1958; Labov and Waletzky, 1967). A few, like the turn in a story, are Sayfield's own call. Each move's study will show under it once it has been checked against the paper.
The way the notes are written comes from feedback research: in studies of learners in schools and universities, feedback about the task and how to do it better helped learning more, on average, than praise (Hattie and Timperley, 2007) or rewards and punishment (Wisniewski and colleagues, 2020). So the notes quote you, suggest one thing to try, and never grade you.
On most interview, exam, work, pitch and explain questions, the notes can use what Sayfield measures: your pace, the 'um' and 'uh' it heard, how soon you got to your point, where you paused and whether a pause ran over a second, and where you restarted just before a figure or a name. Sayfield writes every number itself; the AI never does. No number is a target, and the only comparison is with your own earlier answers. Nothing is said about your voice or the way you say a word, and since fillers like um, euh or ya'ni do real jobs in speech, no note asks you to cut them.
A painful story still gets its notes, opened by a line Sayfield wrote to acknowledge it. The thing to try only adds to the story or joins two parts of it, and no fix or number comes with it. On interview, exam and work questions, the notes follow that question's usual rules.
An AI can get things wrong, so a note can be wrong too. If one misses, tap the thumbs-down beside it and tell Sayfield what's off. On the thing to try, the same sheet lets you pick another move for Take 2.
The AI company that writes the notes gets your words and those measurements as text only. It does not use them to train its models, and deletes them within 30 days unless its safety checks flag them. More on who sees what: Privacy
Small fixes
On interview, exam, work and pitch questions, up to two small fixes can sit under the thing to try. A fix points at a few words that setting would not usually expect, or at one word you used many times, and shows your line again with those words crossed out and, now and then, up to three words to say instead.
A fix is about the setting, never about you or your opinion. Sayfield keeps no list of words you shouldn't say: each word is weighed in its setting, and “Why?” under each fix says what that setting usually expects. Everyday questions, stories and explain answers never get fixes.
Your transcript never changes: it stays the record of what you said. The words a fix points at only get a light underline and a small Fix tag.
If the same sound or word keeps coming up across your answers, Sayfield may offer to keep an eye on it for a week. Say yes, and each new note shows how many times it came up; Stop ends it.
We never score your accent
No grade for you as a speaker, no labels for confidence or emotion. Sayfield describes what was said and how, never the person who said it.
To see what's kept and who can see it, open Privacy