Voice prompts guide each report field and responders answer in natural speech. Captures unit ID, incident type, patient condition, and destination as structured slots, along with free-form narrative dictation. Runs entirely on the device.
The on-device voice guided reporting pipeline runs five Picovoice SDKs in a loop on the responder's device: Porcupine Wake Word, Rhino Speech-to-Intent, Cheetah Streaming Speech-to-Text, Orca Text-to-Speech, and Koala Noise Suppression. Porcupine listens for the wake word. Koala suppresses sirens, traffic, and ambient field noise. Orca speaks each report field aloud. Rhino captures structured slots — unit ID, incident type, patient condition, destination — directly from speech. Cheetah transcribes free-form narrative dictation. Audio never leaves the device.
Ambulance bays, scene perimeters, and field sites are loud. Koala Noise Suppression cleans sirens, traffic, generator hum, and crowd noise before the audio reaches Porcupine, Rhino, and Cheetah. Koala suppresses background noise twice as effectively as RNNoise with the same compute footprint, which leaves headroom for the rest of the pipeline on embedded devices, legacy phones, and rugged tablets from Honeywell or Zebra.
Porcupine Wake Word starts the reporting workflow when the responder says the chosen phrase. Crews can train a branded wake word or always-listening commands, such as “New incident” or “Begin charting” in the Picovoice Console, and deploy them across mobile and embedded devices. Porcupine runs always-on at low CPU and battery cost so the rest of the pipeline only spins up when needed, critical for EMS shifts that run twelve hours on battery.
Rhino Speech-to-Intent captures structured ePCR slots, such as unit ID, incident type, patient condition, transport destination, handoff status, and handoff time, directly from speech. Most voice command systems run a two-step pipeline: speech-to-text produces a transcript, then a separate NLU model parses that transcript for intent. Each step accumulates error and compounds latency. Rhino infers intent and typed slot values directly from audio, holding higher accuracy in noisy environments without hallucinations or compounding errors.
Cheetah Streaming Speech-to-Text transcribes the responder's free-form narrative dictation in real time, including patient details, scene description, treatments administered, and witness statements. Per the open-source real-time transcription benchmark, Cheetah beats Google Cloud STT on word error rate and word emission latency across all tested languages, and outperforms Azure STT on several. It emits words at 590 ms median latency, typically one word behind the speaker, and requires less compute than any other local engine tested. Cheetah accepts custom vocabulary for clinical terminology, drug names, hospital destinations, and proper nouns, which raises accuracy further on the words responders actually say.
Orca Streaming Text-to-Speech reads each report prompt aloud, for example, “What's your unit?”, “What's the patient's condition?”, or “What's the transport destination?”, so the responder never has to look at the screen while driving, scene-managing, or treating a patient. Most high-quality TTS engines require hundreds of megabytes of RAM. Orca uses 29 MB peak memory, 10 to 50 times less than any natural-sounding on-device alternative, which leaves enough headroom to run all five engines on a single rugged tablet without OOM crashes. First-token latency is 130 ms, fast enough that prompts feel conversational rather than robotic.
Electronic Patient Care Reports are the run records EMS crews complete for every patient encounter. Voice ePCR apps capture unit ID, incident type, patient condition, transport destination, and free-form narrative dictation while crews are en route, ready to map to ESO, ImageTrend, ZOLL RescueNet, HealthEMS, or a NEMSIS-compatible schema.
Officers and firefighters document incidents from the scene without typing. Voice reporting apps capture call type, location, person of interest, response time, and handoff status as structured slots, plus witness statements as free-form dictation. Output can map to municipal incident-report formats and CAD-RMS schemas.
Utility crews, telecom field technicians, and oil & gas inspectors capture structured observations hands-free where cell signal drops. Each Rhino context tunes per asset class: pole IDs and circuit numbers for utilities, tower IDs for telecom, and well or pipeline segment IDs for oil & gas.
Insurance claims at incident scenes, environmental surveys, wildlife observations, and humanitarian field reporting need a structured-plus-narrative voice flow that does not depend on a cell tower. The on-device pipeline runs in basements, canyons, and offshore platforms where Wi-Fi and cellular do not reach.
A complete working recipe in Python. Open-source on GitHub. Runs 100% on-device.