Dr Lidea Shahidi is a postdoctoral research associate in the Deep Hearing Lab, part of the Cambridge Hearing Group at the MRC Cognition and Brain Sciences Unit. An engineer by training who finished her PhD in 2022 at Duke University, she now focuses her research on improving speech perception outcomes for individuals with cochlear implants. A cochlear implant is a small, complex electronic device that can help to provide a sense of sound to a person who is profoundly deaf or severely hard-of-hearing (https://www.nidcd.nih.gov/health/cochlear-implants).
In 2024, an award from the Cambridge Language Sciences Incubator Fund (funded through CLS EQUILL-AI project in collaboration with ai@cam) allowed her to lead her first independent project, "Generative Speech Corpora for Ecological Speech Perception Assessment", testing whether an AI-based method for creating speech corpora could modernise the speech tests used in audiology clinics.
Dr Shahidi was in conversation with Shrankhla Pandey, a PhD student at the Department of Psychiatry and the Department of Computer Science and Technology.
What problem were you trying to solve?
If you have a hearing loss and go to a clinic, they can measure how much hearing you've lost. There are also functional consequences of that hearing loss, and speech is one of the main things we use our hearing for day to day. So the clinic runs a speech test: they play you a recorded sentence, you report back the words you heard, and that tells us how much your hearing loss affects your ability to perceive speech.
The trouble is, it isn't a perfect measure. In the UK, one of the most common tests has the same male speaker saying around 200 sentences that were written and recorded back in 1979. The accent has dated, and the content doesn't always reflect what people actually hear every day. Moreover, if you visit the clinic often, you end up hearing the same sentences again and again, and you can likely guess the next word because you've heard it before. So we asked, "Can we do better?"
What does it take to build a speech test and how did you approach it?
Making speech tests is tedious and time-consuming. You have to find someone with the voice you want, put them in a recording booth, write sentences that are phonetically and semantically balanced, record them, do post-processing and calibration so every sentence is roughly the same difficulty and loudness. Then you calibrate the test against real listeners to see how much natural variation to expect.
So the engineer in me looked at it and the question became: we have voice-cloning technology now; can we use it to make the process of building speech tests more efficient, with a faster turnaround, while expanding the representation of people's voices and accents, and the languages used? Ultimately the vision was, and still is, to have a speech test better suited to someone's actual day-to-day listening, so their score reflects their real-world hearing more faithfully.
Why was this the right moment for the project? What made it possible now?
Two things made this the right moment: the technology and the healthcare community were both ready.
First, the technology. The project is very much driven by the recent wave of open-source text-to-speech models. Because their versions are fixed, these models give reproducible results, which is the whole point of research. It matters especially for clinical testing: we want the model to stay the same so that we can track the change in a person's measurement longitudinally. If you extended a test and someone scored higher later, you'd want to know whether they'd actually improved or whether the synthetic speech was just easier to understand. That said, I don't think there'd be this much excitement around open-source models without the commercial ones making the technology more visible. There's a lot of fast-paced work on synthetic speech right now, and most of it uses proprietary models, but that's at odds with open science.
"I could run an experiment today and again a week later and get a different result, with no way to know if it's the model or the measure that changed." [due to the limitation of a proprietary model]
A proprietary model can change without telling you. You don't get a snapshot of it in time; you often can't even say what kind of model it is. The online interfaces have knobs labelled things like "stochasticity" [or randomness - the property that makes a model give slightly different outputs each time] (https://en.wikipedia.org/wiki/Stochastic) and you've no idea what they actually do.
The second thing that made this project successful was the support of the healthcare community. It was well received among audiologists and surgeons when I presented it at the annual meeting of the British Cochlear Implant Group. People found it interesting and were looking forward to seeing more. It was rooted in interdisciplinary collaboration. Early on, we contacted Dr Marina Salorio-Corbetto, who works at the SOUND Laboratory, embedded in the Clinical School (and led by Professor Debi Vickers) and she's been a big help in knowing what would be feasible and what measures we need to take to make this clinically credible and well-vetted.
Tell me about the work itself: what you did and what it showed.
We implemented more than twelve open-source models and evaluated the most promising ones – three of which we reported in detail in a paper accepted at Interspeech 2025, a leading speech-technology conference, and presented in August 2025. Honestly, there are more models every month, and we had to force ourselves to stop and pick a few. Each one takes time and storage just to set up the environment. Comparing them openly and systematically is something our project did well. The project benchmarked these models in an online study with native British English listeners, measuring three things at once: how intelligible the speech was in background noise, its quality, and whether listeners could tell the cloned voice apart from the original. One model, E2, reproduced the original speaker so convincingly that listeners often judged its samples to be more "him" than the real recordings.
The interesting - and slightly concerning - finding is that cloned speech can be more intelligible than the human speech it was cloned from. That benefit even survives vocoding [vocoding strips speech down to a small number of frequency bands, used here to simulate what a cochlear implant user hears] (https://en.wikipedia.org/wiki/Vocoder). From a clinical angle that matters: you'd want to know whether a higher score reflects real improvement or just easier-to-understand speech.
What challenges did you face in creating these new AI-powered speech tests?
Our starting point was to reproduce the existing test, and that meant voice-cloning the person who originally recorded it. This posed an ethical challenge! The original speaker of the recordings gave us his permission to clone his voice; this is invaluable as without him, we couldn't say we'd reproduced this exact test. He still comes to the UK Auditory Science Meetings on occasion, and we were able to present our results to him then. We keep him updated on our progress and he’s been quite supportive and interested to see how the project progresses.
Another challenge was making this test multilingual. That was our goal when we applied for the funding, but the models that claim to be multilingual tend to have poor voice cloning, which gets worse for low-resource languages. Our goal to make a well-represented modern speech test remains, but the technology to make it feasible across multiple languages is not there yet.
How did the Incubator funding call, or CLS in general, help you build new collaborations?
This was my first independent project, so I'm really grateful to Cambridge Language Sciences for making it possible. Once we in the Deep Hearing Lab had the idea, the question was: what are we missing? And it was clinical and linguistic expertise. So I started looking up people in Cambridge, working out what they could bring, and emailing them.
That led directly to working with Dr Alexis Deighton-MacIntyre, a fellow member in the Deep Hearing Lab, who brings deep expertise in speech rhythm and neural tracking [neural tracking is a method that measures how closely the brain's electrical activity follows the rhythm of speech] (https://en.wikipedia.org/wiki/Cortical_entrainment). Her insight into what might make synthetic speech different from human speech has been hugely informative.
As I mentioned earlier, on the clinical side, Dr Marina Salorio-Corbetto at the SOUND Lab and the Emmeline Centre advised on what would be feasible and well-vetted from a clinical perspective.
The project was led by me, under the guidance of my PI, Dr Tobias Goehring.
"Working across disciplines brings in new perspectives. From an engineering perspective, what matters is that the synthetic voices sound natural and human-like, whereas a linguist will be more concerned with accurate phonetic reproduction and a clinician with consistent and reliable speech test outputs.”
What happened after the grant ended?
It opened up a whole research line for me.
"Without it I'd be interested in synthetic speech and probably messing about with models online, but I wouldn't have research questions I could actively explore, like I did here."
The proof-of-concept led to a follow-on grant from the Accelerate-C2D3 call (circa £27,000) with Dr Deighton-MacIntyre, to test how well voice-cloning reproduces spontaneous and conversational speaking styles, and it's now supporting our work towards a synthetic-speech-based clinical assessment. We applied for further funding from the Hearing Industry Research Consortium; that one was unsuccessful, but we're continuing to develop it for longer-term support.
Why was the Incubator Fund the right fit?
It's well suited to starting up a very small, well-encapsulated feasibility project. It allowed me to pursue a new idea that was different from my main research theme and let me expand it.
"This award has been critical for my career success, because it let me pursue a new idea outside my main research theme, and made me a far more competitive applicant for the fellowships I'm now going for."
The Incubator Fund application was short: it let you say "this is feasible" without pinning down every experimental detail in advance, which was a real advantage for me, coming up with a new speech test. The nine-month duration was good; shorter would have been difficult given the preparation needed to implement the models and run behavioural experiments.
"These smaller pots of money are rather rare and can be critical for us producing preliminary evidence for tools and techniques that can useful in hearing clinics and out in the real world."
The fund left a lasting impact beyond me and the research field. It supported a summer student through the Cambridge Mathematics Placement programme, Erdem Baha Topbas, who worked on implementing and evaluating the models. He stayed on as a research assistant for a year, and has since started a PhD in computational linguistics at the University of Zurich, inspired by this work. The fund offered a consistency that was wonderful; the same person carried the technical work across the whole arc of the project.
You mentioned this was the first project where you took the lead. Any advice for those starting their own?
I learned a lot about running a collaboration, including that not everyone has time to contribute, and that as the lead you do have to give people explicit tasks and actually direct them, even when you feel too junior to be dictating to others what to do. I’ve certainly grown in my ability to lead a project and my confidence to pursue exciting scientific ideas.