Unstable call audio
Background noise, volume differences, and connection quality all affect recognition.
Technology & R&D
Putting voice AI into real operations, not just demos.
AnyTaxi's R&D focuses on validating voice processing, order workflows, and dispatch systems within the same operating context. We define the problem and failure modes first, then choose testable technical approaches.
Technical problems
The hard part of voice orders is rarely a single model.
Call audio, spoken addresses, operational workflows, and existing systems are interconnected. Instability in any one stage shows up in order quality and manual workload.
Spoken address variants
Homophones, abbreviations, partial descriptions, and local names mean addresses cannot be matched literally.
Balancing latency and review
Excessive automation amplifies errors; excessive manual work removes the value of process improvement.
Integration cost
Dispatch systems, data formats, and permission models vary. Interface boundaries must be defined first.
Concept architecture
Layered, so every stage can be tested and replaced independently.
Voice entry, voice processing, the operations core, and observability connect through explicit interfaces. This lets us replace voice services, adjust extraction rules, or use human review where a stage is not yet stable.
Disclosure boundaries
This website does not disclose server addresses, credentials, keys, internal admin URLs, customer personal data, or sensitive deployment details.
Voice entry
Inbound callVoice workflow
Voice processing
Speech-to-textIntent & field extraction
Operations core
Order serviceDispatch interfaceHuman review
Data & observability
Operations dataEvent logTest samples
Cross-cutting: access control · logging · monitoring · data minimisation
Engineering approach
Make the system inspectable before scaling it.
Testing, records, and human review create a feedback loop. Without a comparable baseline, it is difficult to know whether a change actually improves the workflow.
Fixed test sets
Use manually labelled call samples to create a repeatable evaluation baseline.
Event records
Record outcomes and failure reasons at each stage to support tracing and debugging.
Human review
Low-confidence or conflicting information goes to human confirmation rather than straight into an order.
Data minimisation
Retain only what the workflow needs, with access control and retention limits.
Evaluation method
Define how to measure before claiming improvement.
We do not publish unverified recognition rates or customer outcomes. The metrics below establish a repeatable baseline; values are published only after confirmation on fixed test sets and field conditions.
Field correctness
Pickup, time, passenger count, and notes are compared against manual annotation.
Human review rate
The rate, reasons, and corrections for low-confidence or conflicting information.
End-to-end latency
Time from call end to a dispatchable order; SLA must be defined with the field.
Order and dispatch reconciliation
Consistency across voice, order, dispatch, and completion states.
Risks and mitigation
Design for failure modes.
Call audio, spoken addresses, latency, and integration can fail. We reduce risk with human review, event records, and staged validation rather than assuming the model is always correct.
Audio and accents
Pre-processing, fixed test sets, and scenario-level observation.
Address variants
Normalisation, candidate matching, low-confidence follow-up, and human review.
Latency and capacity
Staged rollout with end-to-end latency and failure-rate measurement.
System integration
Interface boundaries, sandbox validation, event records, and rollback.
R&D status
Completed, in validation, and in R&D are stated separately.
The status below reflects the October 2026 internal audit and field evidence. It describes inspectable work, not website test results; raw evidence and sensitive data are not published.
-
Telephony and fleet back officeCompleted
Production telephony, human transfer, call records, and the fleet back office have operating records.
Verification：Operating records and internal capability audit
Completed -
Speech and address field validationIn validation
Field use and low-confidence follow-up continue; per-language recognition and order-creation rates are not published.
Verification：Fixed test sets, manual annotation, and field feedback
In validation -
Dispatch and end-to-end reconciliationIn validation
The dispatch engine and back office are used in the field; reconciliation from voice to order to dispatch remains under validation.
Verification：Event records and field reconciliation
In validation -
Voice learning platformIn development
Observation and manual annotation continue; automated candidates are not activated.
Verification：Versioned data and manual annotation workflows
In development -
Multi-fleet and white-labelIn development
Multi-tenant isolation and second-fleet onboarding still require testing and field evidence.
Verification：Isolation tests and second-fleet onboarding records
In development -
Cross-language voice workflowPlanned
Language and scenario priorities will be assessed with field requirements; implementation has not started.
Verification：Field requirements and language priorities
Planned
R&D roadmap
Near-term validation, mid-term productisation, long-term research.
The following describes research direction and work order, not committed delivery dates or revenue targets.
Near term
- Address coverage and low-confidence follow-up
- Human review and exception handling
- End-to-end reconciliation
- Multi-fleet isolation testing
Mid term
- Voice-learning candidates and version governance
- Cross-language and white-label workflows
- Operations data products
Long term
- Multi-agent governance and automated operations
- Industry partnership and ecosystem integration
Work with us
Need technical cooperation or more R&D detail?
If you are evaluating voice workflows for fleets, AI automation, or system integration, get in touch to discuss a verifiable scope.