Ear
A candidate open speech model would turn audio into text with speakers and times and produce tags for dictation commands.
CareCompile speech program
Architecture and evaluation before claims
Vocero separates speech, dictation rules, and meeting analysis so each component can be tested. The architecture, training, and benchmark form a technical plan; the site identifies separately what has already been built.
In development for internal use. Our own model is not trained yet.
01 / VOCERO
A cascade of components
The design avoids assigning to the language model work that belongs to speech recognition or explicit rules.
Rules
A non-AI layer applies punctuation, formatting, and corrections. This layer is built, and its demo version runs in the browser.
Brain
A candidate open base model would turn the transcript into decisions, tasks, entities, and answers with references.
Review
The final output requires human review before it is saved or sent.
02 / VOCERO
Training in stages
Every proposed stage has an evidence gate. No training stage for Vocero's own model has been run.
Baseline first
The plan compares open candidates on a sealed per-country set before selecting models or starting tuning.
Regional tuning next
The ear would be tuned on regional audio mixed with general Spanish. The brain would be tuned on minutes, tasks, and cited answers.
Continued pretraining only if needed
Regional text would be used for continued pretraining only if the baseline reveals a vocabulary gap that a glossary cannot solve.
Controlled serving
The objective is to quantize and serve on CareCompile equipment, behind a switch that is off by default and requires explicit approval to activate.
03 / VOCERO
Vocero-Bench Central America
The benchmark is designed to evaluate each country and separate development from testing. No audio has been recorded, no set sealed, and no model scored.
Speech and dictation
Planned metrics include word error rate, character error rate, punctuation, commands in order, and false activations.
Minutes and answers
Two native reviewers per country would score completeness, accuracy, usefulness, and register, and disagreements would be published.
Invented content
Any name, number, or decision absent from the transcript counts as invented; a report would not claim zero without the corresponding statistical evidence.
04 / VOCERO
Controls for a verifiable comparison
The proposed method aims to prevent training from seeing or reproducing the test set.
Split by speaker and meeting
The design assigns 80% to sealed testing and 20% to public development, with no speakers or meetings shared between them.
Independent custodian
The person holding test audio does not train models; the training team would receive only aggregate results by country.
Fingerprint and contamination
The plan publishes the SHA-256 manifest fingerprint and checks hashes, speaker identity, and thirteen-word overlaps before each run.
Responsible publication
When results exist, they should include benchmark version, date, compared systems, and method. There are no figures to publish today.
Explore the program
Related pages
From a meeting to reviewable minutes
Explore Vocero's design for transcribing meetings and producing minutes with decisions, tasks, and source citations. In development; the demo is synthetic.
Explore the scopeDictation for regional Spanish
Vocero investigates dictation for Salvadoran and Central American Spanish. Try the implemented command rules and explore the regional speech plan.
Explore the scopeA counterpart for Salvadoran voices and texts
CareCompile seeks an institutional counterpart in El Salvador to explore consented speech and text data, limited purpose, and independent evaluation.
Explore the scopeEL SALVADOR / CARECOMPILE
Designate an institutional counterpart in El Salvador
We are looking for an institution willing to discuss regional meetings, voices, and texts. We explain purpose, consent, and governance before recording or receiving data.