For years, the biggest problem in speech therapy has not been knowing what works. Clinicians have known that for decades. The problem has been reach. There are simply not enough speech-language pathologists (SLPs) to see every child who needs one, at the frequency the evidence says they need. Waiting lists stretch for months, school caseloads balloon into the hundreds, and families are sent home with a photocopied word list and the vague instruction to “practise every day.”

That gap (between what good therapy requires and what the system can actually deliver) is exactly where AI speech therapy is starting to take hold. And it is worth looking closely at what these tools do well, where they fall short, and what their arrival tells us about the direction of speech-language pathology and talk therapy more broadly.

What AI speech therapy actually does now

The most tangible example of this shift is a new class of consumer-facing apps built alongside clinicians rather than in place of them. SpeechLP, an AI-driven articulation app that launched at the American Speech-Language-Hearing Association (ASHA) 2025 conference, is a useful case study.

Rather than replacing the therapist, SpeechLP targets the part of therapy that has always been hardest to supervise: the practice that happens between sessions. Children work through game-based articulation exercises, and as they speak, on-device AI listens at the phonetic level and gives real-time feedback: a small, immediate signal of whether a sound was produced correctly. Progress is tracked sound by sound, and by position within a word (initial, medial, final), so a parent or SLP can see whether a child’s trouble with /r/ is easing at the start of words but sticking at the end.

That granularity matters. Carryover (getting a newly learned sound out of the therapy room and into everyday speech) is where a lot of progress is won or lost. A tool that makes daily home practice feel like play, and that quietly measures it, is addressing a genuine clinical bottleneck rather than a marketing one. It is also designed with the guardrails you would want around children’s health data, with HIPAA and COPPA compliance and audio processed on the device.

None of this makes an app a therapist. But it does something a busy clinic struggles to do: it turns “practise at home” from a hopeful instruction into a measured, repeatable routine.

The clinical case: Why frequency and feedback matter

To understand why tools like this are landing now, it helps to think like a clinician about dosage. Speech sound disorders respond to high-frequency, high-repetition practice with accurate feedback. In an ideal world, a child would get many correct repetitions every day, each one confirmed as right or gently corrected.

In the real world, a school-based SLP might see a child for twenty or thirty minutes once or twice a week, split across a small group. Millions of children in the United States have speech sound disorders, and the profession has been openly grappling with caseload pressure and long wait times for services. When surveys of SLPs report widespread caseload overload and multi-month waits, the shortfall is not a matter of clinical skill; it is a matter of hours in the day.

AI closes part of that gap not by being smarter than the therapist, but by being available more often. Real-time feedback on articulation is a narrow, well-defined task: exactly the kind of pattern-recognition problem machine listening is now reasonably good at. Used as a supplement, it multiplies the number of quality repetitions a child gets without adding to the clinician’s hours.

Where AI is taking speech-language pathology next

Articulation apps are the visible tip of a much larger shift. Across the field, AI is quietly moving into the parts of an SLP’s week that have nothing to do with sitting across from a patient.

The most immediate wins are administrative. Documentation, progress notes, and report writing consume an enormous share of a clinician’s time, and generative AI is already being used to draft and structure that paperwork. Automated screeners can flag children who may need a fuller assessment, helping schools and clinics triage rather than assess everyone from scratch. Researchers are even using large language models to generate clinical teaching cases, expanding the material available to train the next generation of SLPs.

Look further out and the ambitions grow. Speech carries an unusually rich signal about health (it engages neurological, motor, respiratory, and vocal systems at once) which is why researchers are building AI tools to detect and monitor conditions from aphasia and stuttering to the early markers of neurological disease. The direction of travel is clear: AI as a diagnostic and monitoring layer that sits underneath the clinician’s judgement, not on top of it.

The talk therapy parallel

The same story is unfolding, more contentiously, in mental health. Talk therapy faces its own version of the access crisis: roughly half of people with a mental health condition receive no treatment at all. That vacuum has pulled in a wave of AI chatbots promising support at any hour.

For a long time the evidence here was thin; plenty of enthusiasm, few rigorous trials. That began to change in 2025, when researchers at Dartmouth published the first randomised controlled trial of a fully generative AI chatbot for mental health treatment, testing a tool called Therabot with just over 200 adults experiencing depression, anxiety, or high risk for an eating disorder. Reviews of cognitive behavioural therapy chatbots have similarly found meaningful short-term reductions in depression and anxiety symptoms.

But the medical read on these results is cautious for good reason. Most trials are short, samples are small, and it remains unclear which ingredient is doing the work: the therapeutic technique, the sense of being heard, or simply having something respond at 2am. Clinicians consistently warn against treating a chatbot as a substitute for a human, particularly in crisis, and there have already been troubling reports of vulnerable users relying on general-purpose AI in ways it was never designed to handle.

The medical bottom line

Put the two fields side by side and the same lesson emerges. AI is strongest where the task is narrow, measurable, and repetitive: counting correct /s/ sounds, drafting a note, delivering a structured CBT exercise. It is weakest, and potentially dangerous, where the task is relational, ambiguous, or high-stakes: reading a child’s frustration, catching the thing a patient is not saying, holding safety in a crisis.

There are also fairness concerns that a medical lens cannot ignore. Speech recognition systems have shown higher error rates for some dialects and racial groups, which means an AI that mis-hears a child’s accent could misjudge their progress. Any responsible deployment needs human oversight, transparency about limits, and validation across the real diversity of the people it serves.

The honest conclusion is neither hype nor dismissal. Tools like SpeechLP point toward a future where AI handles frequency, measurement, and paperwork, and frees clinicians to do the irreplaceable human work: building rapport, exercising judgement, adapting in the moment. In speech-language pathology and in talk therapy alike, the goal should not be an AI that replaces the therapist. It should be an AI that gives the therapist back their time, and gives every patient more of the practice, feedback, and access that good care has always required.

Adam Mulligan, a psychology graduate from the University of Hertfordshire, has a keen interest in the fields of mental health, wellness, and lifestyle.