Practice that meets your voice where it is

Speak with
confidence.

Evidence-based tools for anyone working on how they speak, whether you stutter, are navigating a speech difference, or simply want to be heard more clearly.

See the Workbench See it in action

The problem

Stuttering is measured by hand, one syllable at a time.

To score a single sample, a clinician replays the audio, counts every disfluency, sorts each one by type, times the long ones, and works out the percentages by hand. It is careful, skilled work. It is also the reason objective measurement often gets skipped, and the reason progress between sessions goes unrecorded.

Counting eats the hour

Time spent tallying is time not spent on therapy, planning, or the client in the room.

Samples are small

Because hand-counting is slow, assessment rests on short clips rather than how someone speaks all week.

No shared record

Practice at home and work in the clinic rarely meet in the same set of numbers.

What's inside

Built for daily voices and busy clinics

drag to explore

Fluency model

The core model: detects every disfluent moment, then sorts it into the clinical types.

The SLP Workbench

The model flags disfluencies on real audio; clinicians confirm the calls and keep the record.

Mobile app

One big record button in your pocket, for individuals at home and clinicians mid-session.

Live recording

Pacing and disfluency cues as you speak, so a session can be adjusted while it is happening.

Practice that fits your day

Short guided sessions with real-time feedback, built around techniques clinicians already use.

Direct Zoom integration

Telehealth sessions stream straight into the Workbench, with live pacing cues during the call.

One-click reports

School, IEP, and progress summaries assembled from confirmed labels, ready to send.

Private by design

Encrypted in transit and at rest, anonymized before training, and never sold to anyone.

More models coming

Pacing, clarity, and other speech differences, trained on the confirmed labels built along the way.

For clinicians

The Workbench listens first. You make the call.

Open a recording and the events are already marked: every repetition, prolongation, and block, tagged with a type and a timestamp. Drag a flag to move it, pull its edges to change its span, retype it in one click. The counts and percentages update as you work.

Label as S W A B Intj Rev PR

Every event, on the waveform

Model-tagged repetitions, prolongations, and blocks, editable by hand.

Standard measures, computed

Stutter-like and total disfluency rates, type distribution, speaking rate, weighted severity.

Reports that write themselves

Confirmed labels become progress summaries, school reports, and session notes.

Open the Workbench See it with sample data

For individuals

Therapy is an hour. Speaking is every day.

The phone app is a recorder first: one big button, wherever you are. Order a coffee, run a standup, read a passage. The model listens for the same events a clinician would mark, so what happens outside the clinic finally counts.

Capture in the moment. Real speech in real situations, not drills in isolation.
Measured against you. Your own baseline, not a population average.
You choose who sees it. Recordings stay private from your SLP until you send them.

Start practicing Create an account

For developers

The same models, behind one endpoint.

Teletherapy platforms, EHRs, research groups, and consumer speech apps can send audio and get structured disfluency events back. Every key chooses its own data path, and healthcare workloads stay sealed by default.

Teletherapy EHR & documentation University research Voice products

Read the docs Get an API key

POST /v1/analyze

{ "audio": "<base64>",
  "return": ["events", "measures"] }

→ { "events": [
    { "t": 4.12, "type": "block",
      "dur": 1.4, "conf": 0.86 },
    { "t": 7.90, "type": "sound_rep",
      "dur": 0.6, "conf": 0.74 } ],
  "measures": { "sld_pct": 4.5,
    "speech_rate_wpm": 122 } }

How it works

Two stages, and a clinician at the end of both.

Detection and classification are separate problems, so we treat them separately. Splitting them keeps the hard judgement calls where they belong: with the clinician.

Binary detection

The first model answers one question across the whole recording: is this stretch of speech disfluent or not? A narrow question is a learnable one, and it runs on far more audio than any labelled corpus of types.

Type classification

Only the detected moments go to the second model, which sorts each one into the clinical categories: sound and word repetitions, prolongations, blocks, and the non-stuttering-like disfluencies alongside them.

The clinician confirms

No label is final until a qualified SLP accepts it. Confirmations and corrections both feed back as training signal, so the corpus grows as professionally labelled data, not model output grading itself.

Every confirmed label is attributable to a licensed professional. That is what makes the resulting dataset worth training on, and what makes the output defensible in a clinical record.

Research and evidence

Measured the way the field already measures.

The Workbench computes the measures clinicians are trained on, using the standard disfluency taxonomy rather than a proprietary score: stutter-like and total disfluency rates, the ratio between them, prolongations and blocks as a share of stuttering-like events, speaking rate against age norms, and a weighted severity that maps to the usual bands.

Validation is the work in front of us, not behind it. Nothing here is peer-reviewed yet, and we will not pretend otherwise. What we can share now is the methodology, and what we want is clinicians and researchers to help design the study.

Help design the study Request the methods

The clinical whitepaper is not written yet

Working on validation before the paper.

Planned

What we intend to report

Agreement between model events and clinician labels
Performance by disfluency type, including the rare ones
Behaviour across ages, accents, and recording conditions
Review time with the Workbench versus by hand

Privacy

Straight answers about your voice data.

Voice is identifying, and clinical voice is protected health information. We are early and we will be plain about it: to start, all model inputs are used as training data. Nothing is sold, everything is anonymized, and the separated paths below are what we are building toward.

SLP

Trains models

Client audio and confirmed labels are handled to HIPAA standards and never sold or shared with other clinics. While we are early, they also train the models; clinic-held sealed paths are on the roadmap.

Individual

Trains models

Practice audio is anonymized and used to improve the models. Per-product opt-outs and separated training paths land as the platform matures.

API

Metered

Audio in, structured events out. While we are early, API audio also trains the models; per-key sealed paths are on the roadmap.

Working to HIPAA standards BAA in progress Encrypted in transit and at rest

Market and business model

One model, three ways to sell it.

Roughly one in a hundred adults stutters, and many more live with other speech differences. They are served by speech-language pathologists in private practice, schools, hospitals, and increasingly over video, all of whom document outcomes and all of whom do it by hand today.

Clinics, per seat

Subscription per clinician, sold into private practices and school districts. The Workbench pays for itself in review time.

Land with one SLP, expand across the practice

Individuals, monthly

A consumer subscription that also, with consent, produces the practice audio the models learn from.

Distribution and data in the same motion

API, metered

Usage-based pricing for teletherapy platforms, EHR vendors, and research groups, with a free tier to build against.

Reach clinicians we never sell to directly

The compounding loop: individuals generate speech, clinicians confirm the labels on it, and confirmed labels make the models better for both. Every product makes the next one cheaper to run and harder to copy.

Roadmap

Where we are, and what comes next.

Built Done

The software is standing up

The Workbench, the individual dashboard, the API, and the iOS recorder: model-tagged review with editable flags and the standard measures computed from confirmed labels.

Now In progress

Legal and compliance

Entity, BAA, and the HIPAA posture behind it, plus the data agreements that let clinics put real client audio through the Workbench.

Next Planned

Validation and the clinical whitepaper

Design the study with clinicians, measure agreement between model output and clinician labels by disfluency type, then publish the methods and the misses.

Later Planned

Beyond fluency

The same detect-then-classify architecture applied to other speech differences, on the corpus of professionally confirmed labels built along the way.

Team

Small, and building in the open.

StutterStep is early. The product is being built alongside the clinicians who will use it, and the parts that are not ready yet are marked as such on this page.

Owen Dinsmore

Founder

Building the models, the Workbench, and the case for both.

Help needed

Speech and audio ML, and a clinical lead who has counted disfluencies by hand for a living.

Advisory board, forming

Practising SLPs, fluency researchers, and people who stutter, shaping what gets measured and how.

Three ways to work with us.

If you assess fluency, teach it, research it, or live it, we want your judgement in the loop early.

Speak withconfidence.