Vocero convierte una reunión de gobierno en un acta con decisiones y responsables, que una persona revisa. Nuestro objetivo es que lo haga en el español de El Salvador, en equipos que controlamos.
Vocero turns a government meeting into minutes with decisions and owners, which a person reviews. Our goal is for it to do so in the Spanish of El Salvador, on equipment we control.
Pedimos una sola cosa: que El Salvador designe una contraparte institucional y abramos una conversación sobre voces y textos, bajo el marco ético.
We ask for one thing: that El Salvador designate an institutional counterpart and we open a conversation about voices and texts, under the ethics framework.
Stage: in development, internal use. Our own model is not trained yet. Demos are synthetic.P Vista de presentación: muestra solo lo esencial. Pulse P otra vez para salir.Presentation view: shows only the essentials. Press P again to exit.
A synthetic meeting, processed step by step: the ear recognizes each turn, the brain extracts entities, decisions and tasks, and every point in the minutes cites its source line. Hover over a point in the minutes to see where it comes from.
Synthetic: sample script, not measured performance.
We start with the training plan for the northern regions.
Three sites are ready; the fourth depends on the computer equipment arriving.
The phase one budget is enough, but phase two has to be requested before quarter close.
So Carlos confirms the sites this week and Elena prepares the request for Thursday.
Done, and I will flag it if the equipment is late.
The minutes are assembled when the brain finishes reading the transcript.
{
"synthetic": true,
"meeting": "Planning call, ministry",
"decisions": [
{
"text": "Phase one is covered by the current budget.",
"source_lines": [
3
]
},
{
"text": "Phase two is requested before quarter close.",
"source_lines": [
3
]
}
],
"tasks": [
{
"task": "Confirm the sites",
"owner": "Carlos",
"due": "this week",
"source_lines": [
4
]
},
{
"task": "Prepare the phase two request",
"owner": "Elena",
"due": "Thursday",
"source_lines": [
3,
4
]
}
],
"open_items": [
{
"text": "The fourth site depends on the computer equipment arriving.",
"source_lines": [
2,
5
]
}
],
"entities": [
{
"type": "place",
"text": "northern regions",
"value": "region: north",
"line": 1
},
{
"type": "amount",
"text": "three sites",
"value": "3 sites ready, 1 pending",
"line": 2
},
{
"type": "date",
"text": "quarter close",
"value": "deadline: quarter close",
"line": 3
},
{
"type": "person",
"text": "Carlos",
"value": "owner of the sites",
"line": 4
},
{
"type": "date",
"text": "this week",
"value": "deadline: this week",
"line": 4
},
{
"type": "person",
"text": "Elena",
"value": "owner of the request",
"line": 4
},
{
"type": "date",
"text": "Thursday",
"value": "deadline: Thursday",
"line": 4
}
]
}
Everything in this demo is a synthetic script running in your browser: timings, waveforms and entities illustrate the output format, not model performance.
Type what a person would say aloud, with the spoken commands: "coma", "punto", "nueva línea", "no perdón". You will see how the dictation rules turn it into written text.
Deterministic rule layer. Everything happens in your browser: no microphone, nothing is sent.
This shows only the dictation rules; speech recognition is not trained yet.
Buscamos voces y textos de todo el país, aportados con consentimiento y bajo reglas claras. Lo que se aporte sirve solo para construir y evaluar Vocero.
We are looking for voices and texts from across the country, given with consent and under clear rules. What is contributed is used only to build and evaluate Vocero.
Hoy Vocero no tiene un foso. Tiene un plan para construirlo y herramientas que lo hacen creíble. Nada está entrenado y todavía no hay clientes.
Today Vocero has no moat. It has a plan to build one and tools that make the plan credible. Nothing is trained and there are no customers yet.
Línea discontinua: todavía no existe. Es el diseño de un ciclo, no uno que ya gire. Cada vuelta solo cuenta si cada paso se puede comprobar por fuera.Dashed line: does not exist yet. It is the design of a loop, not one that already turns. Each turn only counts if each step can be checked from outside.
Un foso empieza el día en que la primera grabación con consentimiento entra al registro y alguien que no entrena modelos sella el primer examen.A moat starts the day the first consented recording enters the ledger and someone who does not train models seals the first exam.
Estos principios aplican a los datos, al entrenamiento y al uso de Vocero. Cada uno se desarrollará en un documento completo. El borrador se comparte a solicitud.
These principles apply to the data, the training and the use of Vocero. Each will be set out in a full document. The draft is shared on request.
Nadie aporta su voz sin saber para qué se usa. El consentimiento es escrito, claro y específico.
No one contributes their voice without knowing what it is for. Consent is written, clear and specific.
Full document: in preparationQuienes aportan, y El Salvador, deciden sobre sus datos. No se revenden.
Contributors, and El Salvador, decide about their data. It is never resold.
Full document: in preparationLos datos sirven solo para construir y evaluar Vocero. Ningún otro uso.
The data is used only to build and evaluate Vocero. No other use.
Full document: in preparationPublicamos qué datos usamos, cómo medimos y cuáles son los límites del sistema.
We publish what data we use, how we measure, and the system's limits.
Full document: in preparationNi vigilancia masiva, ni perfilado político, ni identificar personas por la voz sin su consentimiento.
No mass surveillance, no political profiling, no identifying people by their voice without consent.
Full document: in preparationVocero asiste. Una persona revisa lo que se guarda o se envía, y decide.
Vocero assists. A person reviews what is saved or sent, and decides.
Full document: in preparationEl español de El Salvador, con voseo, de todas las regiones, edades y géneros. No hay un único español «correcto».
The Spanish of El Salvador, with voseo, from every region, age and gender. There is no single "correct" Spanish.
Full document: in preparationAuditoría, derecho a retirarse y a borrar, revisión independiente y reporte de incidentes.
Audit, the right to withdraw and delete, independent review, and incident reporting.
Full document: in preparationVocero is in development and used only inside CareCompile. Dictation works today with an engine that is not our own model. Our own model is not trained yet; that is why we publish no accuracy figures, customers or prices: there are none.
We do not claim a rung until we have its evidence.
Last updated: October 6, 2026, version 3.2.0. This board is updated as the evidence changes.
If your institution can designate a counterpart, or would rather see the demo with synthetic data first, tell us. We talk about voices and texts under the ethics framework before anything is recorded.
Contacto: hola@vocero.icuContact: hola@vocero.icu
Vocero es una cascada, no un único modelo de audio de punta a punta. Cada pieza tiene su propia métrica y se puede medir, corregir o reemplazar por separado. Los comandos de dictado los aplica un motor de reglas, no un modelo de lenguaje. Las salidas pasan por una verificación contra la transcripción y, al final, por una persona.
Vocero is a cascade, not a single end-to-end audio model. Each part has its own metric and can be measured, fixed or replaced on its own. Dictation commands are applied by a rule engine, not by a language model. Outputs are checked against the transcript and, at the end, by a person.
Pase el cursor o use Tab sobre un cuadro: qué hace, cómo se mide y en qué estado está.Hover or Tab onto a box: what it does, how it is measured and its status.
Reconocimiento de voz con puntuación automática en reuniones. En dictado, escribe los comandos como etiquetas (<coma>, <nueva_linea>) en lugar de adivinar.
Speech recognition with automatic punctuation in meetings. In dictation, it writes commands as tags (<coma>, <nueva_linea>) instead of guessing.
Un motor de reglas probado: inserta puntuación, borra o corrige, y todo se puede deshacer. «Entró en coma» no es una coma: solo cuenta como comando en una pausa.
A tested rule engine: inserts punctuation, deletes or corrects, and every action can be undone. "Entró en coma" is not a comma: a word counts as a command only at a pause.
Lee la transcripción y devuelve datos estructurados: acta, tareas con responsable y fecha, y respuestas con la marca de tiempo de donde salen.
Reads the transcript and returns structured data: minutes, tasks with owner and due date, and answers with the timestamp they come from.
Lo que no coincide con la transcripción se marca, no se presenta como hecho. Vocero asiste; una persona revisa y decide.
Anything that does not match the transcript is flagged, not shown as fact. Vocero assists; a person reviews and decides.
In dictation mode, the ear writes spoken commands as tags and a rule layer applies them. Hold the button and watch the raw text become final text.
Synthetic and simulated: does not use your microphone.
| Spoken | Effect |
|---|---|
| «coma» | types , |
| «punto» | types . and capitalizes next |
| «nueva línea» | line break |
| «no perdón» | corrects the previous word |
Done. See how the comma, period, new line and correction were applied.
The sample is a synthetic script; the command layer that processes it is real and runs in your browser. A person should review the text before it is saved.
No entrenamos desde cero: eso requiere miles de GPU. Partimos de modelos abiertos y los ajustamos con datos regionales. Ninguna etapa avanza si no pasa su puerta, y no hay fechas fijas: cada etapa espera a la anterior. Hoy ninguna etapa se ha ejecutado.
We do not train from scratch: that takes thousands of GPUs. We start from open models and adapt them with regional data. No stage moves on until it passes its gate, and there are no fixed dates: each stage waits for the one before. Today no stage has been run.
Correr los modelos base sin tocar, y el motor de dictado que funciona hoy, sobre el conjunto sellado. Error por país y puntaje de actas: la cifra que hay que superar.
Run the untouched base models, and the dictation engine running today, on the sealed set. Error per country and minutes scores: the number to beat.
Reunir, revisar licencias, limpiar y separar por país. El conjunto de evaluación se aparta y se sella antes de cualquier entrenamiento.
Collect, check licenses, clean and split by country. The evaluation set is set aside and sealed before any training.
Ajuste fino de un modelo abierto de voz con audio regional, mezclado con español general para que no olvide lo que ya sabe.
Fine-tune an open speech model on regional audio, mixed with general Spanish so it does not forget what it already knows.
Ajuste por instrucciones (QLoRA) para actas, tareas y respuestas con cita; después, ajuste por preferencias contra invenciones. Preentrenamiento continuo solo si hace falta.
Instruction tuning (QLoRA) for minutes, tasks and cited answers; then preference tuning against invented content. Continued pretraining only if needed.
Fusionar los adaptadores, cuantizar y servir en equipos propios, detrás de un interruptor apagado por defecto.
Merge the adapters, quantize, and serve on our own equipment, behind a switch that is off by default.
Los conjuntos abiertos de voz en español vienen sobre todo de España, México y Sudamérica. Hay muy poco audio abierto de Centroamérica, y sin él el oído no mejora donde importa. Por eso la mayor parte del trabajo real es reunir datos regionales con consentimiento.
Open Spanish speech sets come mostly from Spain, Mexico and South America. There is very little open Central American audio, and without it the ear will not improve where it matters. That is why most of the real work is collecting consented regional data.
Seguir entrenando un modelo ya preentrenado con mucho texto sin etiquetar de la región: documentos públicos, noticias, enciclopedias con licencia. El modelo aprende vocabulario, nombres y giros. No le enseña una tarea. Lo haremos solo si la línea base muestra un vacío de vocabulario que el glosario no resuelve.
Keep training an already pretrained model on a large amount of unlabeled regional text: public documents, news, licensed encyclopedias. The model learns vocabulary, names and turns of phrase. It does not learn a task. We will do it only if the baseline shows a vocabulary gap that the glossary does not fix.
Entrenar con ejemplos etiquetados de la tarea: audio con su transcripción para el oído; transcripción con su acta revisada para el cerebro. Con LoRA solo se entrenan unas matrices pequeñas añadidas al modelo; por eso cabe en una GPU de 24 GB. Un ajuste completo de 8B necesitaría unos 128 GB.
Train on labeled examples of the task: audio with its transcript for the ear; a transcript with its reviewed minutes for the brain. With LoRA only small added matrices are trained, which is why it fits on a 24 GB GPU. A full fine-tune of an 8B model would need about 128 GB.
| Data | What for | Amount | Possible source | Rule |
|---|---|---|---|---|
| Training audio with transcripts | Adapt the ear to Salvadoran speech | ~20 hplanning estimate | Consenting speakers: scripted reading and spontaneous speech; every region, age and gender | Written consent stating the audio trains a model |
| Sealed test audio | Measure without cheating | ~3 hplanning estimate | Speakers different from the training speakers | Never used for training; held by a custodian |
| Dictation clips | Commands, numbers, dates and regional names | to be setdepends on the script | Scripts we write, read by consenting speakers | Same consent as the training audio |
| Regional text | Vocabulary and names for the brain | to be setonly if the baseline needs it | Public documents and news archives whose license allows it | License note per source; no scraping of sites that forbid it |
| Meeting-style examples | Teach minutes, tasks and answers | thousandsfor the whole program | Synthetic meetings we write; real ones only with consent | Labelled synthetic; a person approves each example |
| General Spanish | Keep the ear from forgetting what it knows | existingopen sets | Open Spanish speech sets | License reviewed and recorded before use |
Hours are planning estimates, not commitments or data already collected. No audio has been recorded for this program yet.
Un banco de pruebas diseñado para medir dictado, transcripción, actas y respuestas por país, calificado a ciegas y con intervalos de confianza. Está en diseño: no se ha grabado audio, no se ha sellado el conjunto y no se ha calificado ningún modelo.
A benchmark designed to measure dictation, transcription, minutes and answers per country, scored blind and with confidence intervals. It is in design: no audio has been recorded, no set has been sealed and no model has been scored.
Substitutions, deletions and insertions over reference words, at corpus level. Published with and without accents.
The same formula over characters. Useful with proper names and local words.
Commas, periods, line breaks, and the opening marks ¿ and ¡ as classes of their own.
Command accuracy in order, false activations, and commands typed out instead of executed.
Two native reviewers per country, blind: completeness, correctness, usefulness and register. Disagreements are published.
Any name, number or decision not in the transcript. We never say "zero".
| Country | WER | CER | Punct. F1 | Commands | Minutes rubric | Hallucinations |
|---|---|---|---|---|---|---|
| SVEl Salvador | not measured | not measured | not measured | not measured | not measured | not measured |
| GTGuatemala | not measured | not measured | not measured | not measured | not measured | not measured |
| HNHonduras | not measured | not measured | not measured | not measured | not measured | not measured |
| NINicaragua | not measured | not measured | not measured | not measured | not measured | not measured |
| CRCosta Rica | not measured | not measured | not measured | not measured | not measured | not measured |
| PAPanama | not measured | not measured | not measured | not measured | not measured | not measured |
Cells are empty on purpose: no model has been measured yet. When there are results, they will be published with the benchmark version, the date, the list of systems compared and the method.
Entrenamos y servimos en equipos propios de CareCompile. Las cifras de memoria son objetivos y estimaciones aritméticas, no mediciones; la etapa 0 incluye una prueba corta para confirmarlas antes de cualquier trabajo largo.
We train and serve on CareCompile's own equipment. The memory figures are targets and arithmetic estimates, not measurements; stage 0 includes a short dry run to confirm them before any long job.
QLoRA tuning of a ~8B model and fine-tuning of the ear. Estimate for a typical run: about 12 to 15 GB of 24.
Target: brain at 5 bits with an 8-bit cache, next to the ear, on one card. It fits, but only just: about 10 GB of 12.
Chosen from the baseline, per country, on our sealed set. No candidate has any results from us yet.