Here is a scene that plays out in clinics every week. A doctor tries an AI scribe. The notes come back wrong: words that were never said, a drug that was never mentioned, a line claiming the patient denied symptoms nobody asked about. The doctor decides the scribe is not good enough and goes back to typing.
Most of the time, the scribe was not the problem. The audio was.
By Jhonatan Bringas Dimitriades, MD & Diana van Stijn, MD PhD | Augmented by Keikku AI
The model gets all the attention, but the capture does the real work. And when the capture is poor, the model takes the blame for it.
Why a model invents words
A speech model does not hear like a person. It takes a stream of audio and produces the most likely sequence of words for it. When the audio is clean, that sequence is usually right. When the audio is degraded, noisy, compressed, or recorded across a room on a poor microphone, the model still gives a confident answer. It has no way to say "I did not catch that." It guesses. And a fluent guess reads exactly like a fact.
This is not a rare glitch. Independent research on widely used speech models found that around one percent of transcriptions contain whole phrases that were never spoken, and nearly forty percent of those were harmful rather than harmless. The fabrications cluster around silences and background noise, and they hit hardest the patients whose speech is already hardest to capture. The model invents most where the signal is weakest, and the patients most affected are often the ones who need accuracy the most.
The reverse is also true. Clean up the audio before the model sees it, and both the fabrications and the error rate drop. The output is only as good as the signal it was handed.
The microphone is the first link, not the whole chain
Clinicians who have made AI scribes work all learn the same lesson: check the microphone first. A cheap webcam mic, or a laptop quietly defaulting to one, is the most common reason a good scribe produces nonsense. Swap it for a proper microphone and most problems vanish.
But "buy a better microphone" is where the conversation usually stops, and it should not. The microphone is only the first place the signal can be lost. Between a patient's voice and the finished note, the audio passes through a chain, and every link can throw away detail that nothing downstream can recover:
- Capture. A basic mic starts you with a weak signal. What is not captured cleanly here is gone for good.
- Transport. Over Bluetooth or a phone's call pipeline, the audio is squeezed through a codec built to sound fine to the human ear, not to preserve a faithful signal for a machine. It permanently discards data.
- The cloud hop. Sending audio off the device to be processed often adds another round of compression before the model even runs.
Each stage is one-way. You cannot rebuild later what was lost at the start. A better mic fixes the first link. If the next two are leaking, you are still handing the model a degraded signal and hoping it guesses well.
What radiology already figured out
Radiology settled this years ago. There, lossless is the rule, including for AI, and it is not up for debate. You cannot know in advance which detail the model will need, and a compression artifact can look like disease or hide it. So you protect the raw signal at the source and keep it intact the whole way through.
Audio in medicine deserves the same discipline. Most ambient scribes do not give it that. They simply inherit whatever lossy chain the nearest phone and mic happen to provide.
Owning the signal
The real answer is not a better accessory on a generic pipeline. It is owning the signal end to end: building the capture, the processing, and the path to the model as one system, so the signal is protected from the first link rather than patched at the last.
In practice that means three things. Capture built for the real problem, a room with several voices, distance, and noise, not a mic tuned for a phone at your ear. Processing done close to the source, so the signal is not degraded in transit. And control of the codec itself, because the codec is one of the biggest and most overlooked points of loss in the whole chain.
This is a much higher bar than shipping a branded microphone. Putting a logo on an off-the-shelf module gives you that module's mic, its wireless stack, and its codec, problems and all. Real ownership means building those layers yourself, on purpose, for the job. It is hard, which is why few do it. But it is the only way to fix the cause instead of the symptom.
What this means for choosing a scribe
When evaluating an AI scribe, the instinct is to scrutinize the model. The model matters. But the question that best predicts real-world performance is the one asked least: what happens to the audio before the model ever sees it?
A scribe that captures cleanly and protects the signal will beat a more famous model fed a degraded stream, every time. Fabrications are not just a flaw in the model. They are a property of the signal the model was given. Fix the signal, and you remove the conditions that make it invent.
Garbage in, hallucination out. The reverse is the whole point: protect the signal at the source, and you have done more for the quality and safety of the note than anything downstream can undo.
Sources
The research referenced in this piece is publicly available:
-
Koenecke, A., Choi, A.S.G., Mei, K.X., Schellmann, H., and Sloane, M. "Careless Whisper: Speech-to-Text Hallucination Harms." ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2024. Peer-reviewed. Found that roughly 1% of transcriptions contained entirely fabricated phrases absent from the source audio, that a meaningful share of those carried real-world harm, and that fabrications cluster around silence and non-speech, falling hardest on speakers whose voices are harder to capture. The authors note model providers have since reduced these rates, but the underlying mechanism remains.
-
Noel, M. "Why the Audio Compression Format Impacts the Speech to Text Transcription Accuracy." IBM Watson Speech Services, 2021. An engineering analysis (not peer-reviewed) comparing word error rates across audio formats. Found that lossy compression which sounds identical to the human ear measurably degrades machine transcription accuracy, because decompression introduces artifacts the model then has to interpret.
-
Research on speech recognition and non-speech audio has shown that cleaning the input before transcription, by removing non-speech segments, reduces both hallucination rates and word error rates, reinforcing that the quality of the captured signal, not only the model, governs the quality of the output.




