Hello! I’ve been training speaker diarization models using Common Voice data and noticed an opportunity for improvement: currently, client_id values appear to be hashed differently across the datasets. Scripted Speech uses 128-char hashes, while Spontaneous Speech uses 32-char. Because the hashing schemes differ, there is no identifier overlap between the two datasets, even when the same person contributes to both.
Having recordings from the same speaker across both formal reading and spontaneous conversation, linked by a consistent client_id, would greatly increase the scientific utility of Common Voice for speech AI research, enabling use cases similar to cross-condition benchmarks in NIST SRE and VoxCeleb.
Would you consider aligning speaker IDs across Scripted and Spontaneous Speech releases in future versions? Thanks!
Yes, they are different, because SPS (Spontaneous Speech) is constructed independently without user database connection to SCS (Scripted Speech) - where the user / profile data lives. Even in the first phases it collected its own independent user database. Also, there are no “client_id” in SPS (uses numeric id cols) thus the bundler algorithm is also different.
Would you consider …
Yes we do, actually did as part of our upcoming v2.0 preparation which combines the two independent systems (among other goodies). But we can only match users after the point where the authentication system gets shared.
I’d like to learn about your application more - tomorrow’s meeting?
Nice! This will be a major upgrade, especially for training speaker diarization models.
I spoke more about this during the meeting, but just so everyone here knows: if person A has recordings in both the Scripted Speech and Spontaneous Speech datasets, since the recordings don’t share the same “speaker ID”, when you’re training an AI model, they will be treated as if they came from two different people. If we have the same speaker ID, the model can learn how the same person behaves in different situations (reading and speaking spontaneously).