Accessibility of voice interfaces: what Spanish Law 11/2023 requires

Abstract

Law 11/2023 imposes on the products within its scope three requirements that govern the design of a voice interface: delivery of functions through more than one sensory channel, alternatives to speech and voice operation where the product uses speech, and an alternative to biometric identification and control. The requirements have been enforceable since June 28, 2025, and Royal Decree 143/2026 designated the body that receives complaints as of February 28, 2026. The available evidence describes an asymmetry: in a usability study of assistive robotic arms, 39 of 40 participants operated the voice interface while only 11 of the 20 with neurological impairment operated the joystick; in an evaluation of eight commercial recognition systems, severe dysarthria exceeds a 49 % word error rate in all of them. Whether a given product is covered depends on the closed list in article 2, which does not by itself include every product operated by speech.

Introduction

Design work on a conversational product treats voice as an input and output channel, and judges its quality by recognition accuracy and response latency. The Spanish regulatory framework introduces a different condition: the availability of a route that passes through neither speaking nor hearing.

Law 11/2023, of May 8, transposes Directive (EU) 2019/882 on the accessibility requirements for products and services. Its technical requirements appear in Annex I, structured so that those applying to products are separated from those applying to services, and its enforcement regime was completed in 2026 with the creation of the surveillance and coordination body.

Scope of the law and date of enforceability

Article 2 delimits the scope through a closed list. Products include general purpose consumer computer hardware and its operating systems, self-service terminals, consumer terminal equipment with interactive computing capability used for electronic communications services or for accessing audiovisual media services, and e-readers. Services provided to consumers include electronic communications, the listed elements of transport services, banking services, e-books, e-commerce services, certain elements of utility supply and travel agency services, and social networks.

One consequence of that delimitation governs everything else. A wellbeing product operated by speech is not covered because it is operated by speech: it is covered when it fits one of the categories in article 2. The usual routes in for a consumer product are the e-commerce service through which it is contracted, and its character as consumer terminal equipment when the product is used to access electronic communications services.

The indirect route has the same practical effect. Section III of Annex I, applicable to all services, states in point a) that provision shall be carried out by “ensuring the accessibility of the products used in the provision of the service in accordance with section I of this annex.” A covered service therefore pulls the section I product requirements onto the devices and applications used to deliver it.

The requirements have been enforceable since June 28, 2025, with an extension to June 28, 2027 for emergency communications to 112. Article 30 refers the penalty regime to sector legislation and, on a supplementary basis, to title III of the consolidated text approved by Royal Legislative Decree 1/2013, whose article 83.1 sets fines “from a minimum of 301 euros up to a maximum of 1,000,000 euros,” with ceilings of 30,000 euros for minor breaches and 90,000 for serious ones.

Royal Decree 143/2026, published on February 27, 2026 and in force the following day, created the technical support and coordination unit for the market surveillance authorities, assigned to the Directorate-General for the Rights of Persons with Disabilities. Its functions include acting as a point of information and communication with the public, providing adequate means for receiving complaints and claims, and acting as the surveillance authority in areas where none has been designated.

The requirements that reach a voice interface

Item 2 of section I of Annex I, titled “User interface and functionality design,” contains three points that describe the problem of a product governed by voice.

Point a) requires that, where the product provides communication, operation, information, control, and orientation functions, it shall do so “through more than one sensory channel, which shall include providing alternatives to visual, auditory, spoken, and tactile communication.”

Point b) addresses speech specifically: “where the product uses speech, it shall provide alternatives to speech and voice operation for communication, operation, control, and orientation.” The requirement covers all four functions, so an alternative that covers consultation and leaves control in the hands of voice satisfies it only in part.

Point l) adds that “the product shall provide an alternative to biometric identification and control.” A voiceprint used as an access method is therefore subject to the existence of another route.

A timing condition applies to the technical route for establishing compliance. Version V4.1.0 of standard EN 301 549, of June 2026, is a final draft submitted to the vote phase of the ETSI approval procedure. Conformity with it grants no presumption of conformity while the reference remains unpublished in the Official Journal of the European Union.

Evidence on voice as a route to access

A usability study published in the Journal of NeuroEngineering and Rehabilitation in February 2026 compared voice control of an assistive robotic arm with joystick control in 20 healthy people and 20 with upper-limb impairment of neurological origin, across three activities of daily living. Every participant except one in the impaired group completed the tasks with the voice interface, while 11 of the 20 in that group were able to operate the joystick. Recognition reached 87 % accuracy and the system usability scale fell within the band described as “Good.”

A single-case study published in Frontiers in Rehabilitation Sciences in 2026 evaluated a recognition system trained on the voice of one speaker, a 34-year-old woman with severe dysarthria and global aphasia fifteen years after a stroke. With 1,120 training utterances across 13 target words and 936 evaluation utterances, the personalized system reached 72.65 % word accuracy, above the 56.75 % mean (SD 12.91) of twelve rehabilitation professionals familiar with the patient. The thirteen-word vocabulary and the single-case design confine the result to a proof of concept.

Evidence on voice as a barrier

Evaluation of commercial systems on atypical speech describes the reverse. A December 2025 study measured, without prior adaptation, the performance of eight systems: four conventional recognizers (AssemblyAI, Whisper large-v3, Deepgram Nova-3, and Nova-3 Medical) and four based on multimodal models (GPT-4o, GPT-4o Mini, Gemini 2.5 Pro, and Gemini 2.5 Flash). In severe dysarthria, the word error rate exceeds 49 % in all of them. The largest improvement observed, 7.36 percentage points in GPT-4o, occurs on that baseline.

Stated preference among those who depend on non-visual channels does not support voice as a substitute either. A usability study published in Disability and Rehabilitation: Assistive Technology in 2026, covering a physical activity app used by blind and low-vision people, recorded mixed preferences between voice and text input, and low engagement with the conversation feature because its responses were generic. Its authors conclude that voice interfaces offer accessibility benefits and should complement screen reader interaction.

Criteria for a product operated by speech

Four verifiable criteria follow from the requirements and the evidence, to be applied before a conversational product is published.

Functional parity of the alternative: every function available by voice exists through another route, checked function by function, including control and access functions and not only consultation.

Reviewability of the dialogue: the content of the interaction remains available as consultable text, so that a person who cannot hear the response can read it and a person who was not understood can check what was transcribed.

Declaration of performance by population: the recognition error rate is published broken down by the speech conditions that degrade it, rather than as an aggregate figure that averages them with typical speech.

An alternative to biometric access: where the product identifies a person by voice, another route of identification exists that does not depend on biometrics.

Limitations of the available evidence

The robotic arm study and the physical activity app study are usability studies with small samples and no randomized comparison group. They describe feasibility and stated preference, with no basis for estimating effects.

The personalized recognition study is a single case with a closed thirteen-word vocabulary. Its result gives no ground for expecting the same performance in spontaneous speech or in other people.

The evaluation of the eight commercial systems was carried out without speaker adaptation. It describes baseline performance rather than what is achievable after personalized training, which the other two studies suggest is higher without quantifying it for this population.

Whether a specific product is covered calls for its own analysis of fit within article 2. The Annex I requirements cited here belong to the products section, and their application to a service occurs through the cross-reference in section III.

Conclusions

Law 11/2023 turns the alternative to speech into an enforceable obligation for the products within its scope, with a body receiving complaints in operation since February 2026 and a penalty ceiling of one million euros through the supplementary route. The technical standard that would grant presumption of conformity remains unharmonized. The available evidence supports two statements at once: for part of the population voice is the only route of operation that works, and for another part the error rate disqualifies it as the sole route. The design consequence matches the legal text, which requires the alternative regardless of recognition quality.

Open research lines: Noor Program and Somia Within

Accessibility is a cross-cutting criterion of yeshcube, whose documented mission is to turn validated scientific knowledge into accessible, safe, and ethical technology for human development. The ERL scale includes accessibility among the dimensions that place a solution at its transfer level.

Noor Program is the line devoted to conversational voice capabilities in underrepresented languages. Its question about which speech falls outside a recognition system coincides with the one these data raise for atypical speech. The program is open and its work has not begun.

Somia Within is the gateway through which a conversational integration enters the architecture, and the point at which an accessibility criterion can be required of a third party. The Audio-first approach documents the reasons for and the limits of putting the sound channel first.

Collaborating on accessibility evaluation of voice interfaces

These lines are developed within Allies, the scientific collaboration system of yeshcube, with four partner types and three principles: value for value, traceability, and independence. No partner can veto a publication.

Measuring recognition performance by speech condition requires corpora held by clinical and advocacy organizations. The line is relevant to speech therapy and rehabilitation services that record dysarthric or aphasic speech, to organizations of people with disabilities able to test the usefulness of a declared alternative, and to research teams working on atypical speech recognition.

Discover our Allies program →

Fill in our Contact Form →

References

• Spain. Ley 11/2023, de 8 de mayo, de trasposición de Directivas de la Unión Europea. Annex I, sections I and III; articles 2 and 30. Consolidated text.
• Spain. Real Decreto 143/2026. Creation of the technical support and coordination unit for the market surveillance authorities. In force since February 28, 2026.
• Spain. Real Decreto Legislativo 1/2013. Article 83, penalty amounts.
• ETSI. Final draft EN 301 549 V4.1.0 (2026-06). Final draft in the vote phase; not harmonized.
• Journal of NeuroEngineering and Rehabilitation. “Testing the usability of a voice control system for assistive robotic arms in people with neurological conditions”. 2026. 20 healthy people and 20 with neurological impairment.
• Frontiers in Rehabilitation Sciences. “A speaker-dependent Voice-Input Voice-Output Communication Aid for severe dysarthria and global aphasia”. 2026. Single case, 13 target words.
• Alsayegh, A. and Masood, T. “Zero-Shot Recognition of Dysarthric Speech Using Commercial Automatic Speech Recognition and Multimodal Large Language Models”. arXiv:2512.17474, December 2025. Eight commercial systems.
• Disability and Rehabilitation: Assistive Technology. “Managing physical activity through voice: a usability study of the PCHA app with blind and low-vision individuals”. 2026.