Our pipeline
One recording, three streams, one multi-stream model
Each stream reads different properties of the same recording. Together they return one reading on the clinical scale the study already uses.
What each stream reads
Three views of the same voice
The task decides which streams contribute. Sustained vowels and syllable repetition feed streams 1 and 2; picture description and conversation feed all three.
Stream 1
Deep-learning embeddings
Learned representations of the whole recording, capturing patterns of voice and speech beyond hand-crafted measures.
Stream 2
Acoustic-linguistic features
Interpretable measures of voice and timing: pitch, voice quality, articulation, pauses and speech rate.
Stream 3
Cognitive-linguistic features
What is said and how it is built: word choice, information content, sentence structure and fluency.
Why three
One reading, three supports
Explainable
Traceable to its features
Each reading traces to the features that drove it, for study teams and medical review.
Robust
Limits drift when processing changes
The streams respond differently to the recording chain, which limits drift in the multi-stream reading.
On the scale
Reported where the study looks
The multi-stream reading lands on the clinical scale the protocol already runs.
Tested
Steadier when the phone’s audio changes
Same Parkinson's recordings, three noise-cancellation methods: the kind of change a phone update brings. The multi-stream reading stayed the most consistent.
Reproducibility across three processing versionsICC, 0 to 1; higher is steadier
Cephalgo multi-stream0.75
Whisper0.65
Handcrafted features0.51
HuBERT0.50
Wav2Vec20.40
Contact
Where would speech sit in your next protocol?
Share your therapeutic area and current studies.
Get in touch →