Custom AI Voice Control for People with Disabilities

Custom Voice Control for People with Disabilities: Architecture and Implementation

Imagine a user with cerebral palsy trying to open a bank statement on a mobile app. Each tap requires a minute of effort. The built-in voice assistant doesn't understand the command "show transactions for March"—recognition fails against the background noise of a TV. Familiar? Dozens of such cases exist. Off-the-shelf STT models achieve only 60–70% accuracy with noise above 40 dB, and latency over 2 seconds kills the UX. For people with limited mobility, every second of waiting is a loss of focus.

Our team builds custom AI voice control systems that solve these problems. We have over 5 years of experience in accessibility and NLP, with 30+ projects delivered for government and commercial customers. Our solutions are based on the latest research, including WCAG recommendations and the EN 301 549 standard.

Voice control is the primary input method for people with musculoskeletal disorders, visual impairments, and elderly users with cognitive challenges. We integrate the system with any interface—from web apps to native desktop software.

Why Off-the-Shelf Voice Assistants Don't Work for People with Disabilities

Standard STT systems (Siri, Alice) are not designed for the specific needs of users with disabilities: they don't adapt to accents or speech impairments, don't offer flexible timeout control, and don't integrate with screen readers. Moreover, latency of 2 seconds or more makes conversations unnatural. In noisy environments, accuracy drops to 60–70%, and sensitive data is sent to the cloud.

How We Build the Recognition Pipeline

The foundation is a pipeline: audio stream → VAD → STT (Whisper) → command classifier (LLM) → executor → TTS feedback. Below are the key components in Python.

from faster_whisper import WhisperModel
from openai import AsyncOpenAI
import asyncio
import pyaudio
import numpy as np

class AccessibilityVoiceController:
    def __init__(self, app_commands: dict):
        self.stt = WhisperModel("base", device="cuda", compute_type="int8")
        self.llm = AsyncOpenAI()
        self.commands = app_commands  # {"open profile": handler_fn, ...}
        self.wake_word = "assistant"

    async def listen_and_execute(self):
        audio_stream = self._open_mic_stream()

        while True:
            audio_chunk = audio_stream.read(frames=16000 * 3)  # 3 seconds
            audio_np = np.frombuffer(audio_chunk, dtype=np.int16).astype(np.float32) / 32768.0

            segments, _ = self.stt.transcribe(audio_np, language="en", vad_filter=True)
            text = " ".join(s.text for s in segments).strip().lower()

            if not text or self.wake_word not in text:
                continue

            command_text = text.split(self.wake_word, 1)[-1].strip()
            await self.process_command(command_text)

    async def process_command(self, text: str):
        # Exact match
        for cmd, handler in self.commands.items():
            if cmd in text:
                await handler()
                await self.speak_feedback(f"Executing: {cmd}")
                return

        # Fuzzy intent classification via LLM
        intent = await self.classify_intent_with_llm(text)
        if intent and intent in self.commands:
            await self.commands[intent]()
            await self.speak_feedback(f"Understood, executing")
        else:
            await self.speak_feedback("I didn't understand the command. Please repeat.")

    async def classify_intent_with_llm(self, text: str) -> str | None:
        available = list(self.commands.keys())
        response = await self.llm.chat.completions.create(
            model="gpt-4o-mini",
            messages=[{
                "role": "system",
                "content": f"Determine which command the user's phrase corresponds to. Available commands: {available}. Return only the command name or 'null'."
            }, {
                "role": "user",
                "content": text
            }]
        )
        result = response.choices[0].message.content.strip()
        return result if result != "null" else None

TTS Feedback

Voice feedback is mandatory—the user must hear confirmation. We use Edge TTS (free, low latency). Code:

import edge_tts
import tempfile
import pygame

async def speak_feedback(text: str, voice: str = "en-US-ChristopherNeural"):
    """Speak system feedback using Edge TTS (free)"""
    tts = edge_tts.Communicate(text=text, voice=voice, rate="+10%")

    with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as f:
        await tts.save(f.name)
        pygame.mixer.music.load(f.name)
        pygame.mixer.music.play()
        while pygame.mixer.music.get_busy():
            await asyncio.sleep(0.1)

Web Interface Navigation

For web applications, we use Playwright: the system emulates user actions via browser commands. Example mapping:

# Commands for web navigation via Playwright/Selenium
class WebAccessibilityCommands:
    COMMAND_MAP = {
        "go to profile": lambda p: p.goto("/profile"),
        "open settings": lambda p: p.goto("/settings"),
        "increase font size": lambda p: p.evaluate("document.documentElement.style.fontSize = '120%'"),
        "decrease font size": lambda p: p.evaluate("document.documentElement.style.fontSize = '90%'"),
        "click save button": lambda p: p.click("button:has-text('Save')"),
        "scroll down": lambda p: p.keyboard.press("End"),
        "read page": lambda p: read_page_content(p),
        "fill name field": fill_name_field,
    }

Screen Reader Compatibility

Voice control complements (not replaces) screen readers. Integration via ARIA live regions is mandatory for WCAG 2.1 compliance. Code:

<!-- Voice command status for screen reader -->
<div
    id="voice-status"
    role="status"
    aria-live="polite"
    aria-atomic="true"
    class="sr-only"
>
    <!-- JS inserts: "Command executed: open profile" -->
</div>

<!-- Visual listening indicator -->
<button
    id="voice-toggle"
    aria-label="Voice control"
    aria-pressed="false"
>
    <span class="mic-icon" aria-hidden="true"></span>
    <span class="sr-only">Activate voice control</span>
</button>

Case Study: Voice Control for a Government Services Portal

For one regional portal, we implemented a voice navigation system. Users with visual impairments could fully control the portal without keyboard or mouse. The main challenges were user accents (southern dialect) and the need for action confirmation. We solved them by fine-tuning Whisper on 1000 hours of regional speech and implementing two-step confirmation for critical actions. Results: 96% recognition accuracy, 1.2-second command execution time. The system passed a WCAG 2.1 AA audit.

How We Adapt the System to Individual Needs

For users with speech impairments, we increase the timeout to 10 seconds and add repetition. For accents or dialects, we fine-tune Whisper on the customer's data. For slow speech, we lower the VAD threshold. For cognitive impairments, we use simple single-word commands and voice prompts. In noisy environments, we apply DeepFilterNet before STT, boosting accuracy to 95%.

Testing with Real Users

We involve a focus group of 10–15 people with various types of disabilities. Each scenario is tested on three devices: laptop, tablet, and smartphone. We collect metrics: precision, recall, user satisfaction score. We iteratively refine the model and interface.

What's Included

• Audit of the current interface for accessibility issues.
• Selection of STT and TTS models for your scenario.
• Development of recognition pipeline + LLM classification.
• Integration with frontend (React, Vue, plain HTML).
• Configuration of wake words and user profiles.
• Testing with a focus group of users with disabilities.
• Documentation and training for your team.

Why Our Solution Outperforms Standard Assistants

Our solution is 40% more accurate than standard STT systems under noise, and LLM classification reduces false positives by 2x. Computing resource savings up to 80% thanks to caching of frequent commands.

Ready to Discuss Your Project? Contact us to get a consultation within a day. Order a demo version for testing on your data.

We guarantee that the final solution will pass a WCAG 2.1 AA audit. Our team's experience is confirmed by 30+ successful projects and certification in accessibility.

Speech Recognition and Synthesis: ASR, TTS, Voice Cloning

We tackled a client's challenge: transcribe 40,000 hours of call center recordings in a week. Their existing cloud ASR (Google Speech-to-Text) yielded a WER of 28% on industry-specific vocabulary and cost $0.006 per minute — prohibitively expensive at that volume. The goal was to reduce WER below 10% and switch to self-hosted inference. After deploying a custom pipeline based on Whisper with fine-tuning and faster-whisper inference, the client saved $12,000 per month and achieved a WER of 7.3%.

How does speech recognition ASR handle noisy call center recordings?

The most common issue is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec. By applying loudnorm preprocessing and fine-tuning on 200 hours of labeled data, we consistently cut WER by a factor of 3.

Typical problems we encounter

WER does not converge to the desired metric. Often the culprit is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec.

Diarization fails with more than two speakers. pyannote/speaker-diarization-3.1 works stably for 2–3 speakers, but DER (Diarization Error Rate) increases from 6% to 18–22% with 5+ conference participants. The problem worsens with overlapping speech; by default min_duration_on=0.1 cuts short interjections. We mitigate this with voice-activity detection (VAD) fine-tuning and a custom overlap-handling module.

Voice cloning — latency vs. quality. XTTS v2 (Coqui) delivers natural voice, but during streaming generation stream_chunk_size=20 the first audio chunk arrives after 1.4–2.0 seconds — unacceptable for interactive scenarios. StyleTTS2 and Kokoro are faster but require careful preparation of reference audio.

How do we solve it in practice?

The basic stack for a production pipeline:

• ASR: openai/whisper-large-v3 or faster-whisper (CTranslate2 backend, 4× speed vs original)
• Diarization: pyannote.audio 3.x + integration via whisperx for word-level alignment
• TTS: XTTS v2 for quality, Edge-TTS or Silero for low latency
• Cloning: XTTS v2 (3–6 s reference audio) or OpenVoice v2

A typical call center pipeline: audio from Kafka queue → ffmpeg -af loudnorm normalization to -23 LUFS → faster-whisper with beam_size=5, vad_filter=True → pyannote diarization → post-processing (punctuation via deepmultilingualpunctuation) → write to PostgreSQL with timestamps.

Case study from our practice. A fintech company with 12,000 calls per day. Initial WER on Russian with banking vocabulary — 22% (Google STT). After fine-tuning whisper-medium on 200 hours of labeled recordings via Hugging Face transformers + Seq2SeqTrainer with learning_rate=1e-5, warmup_steps=500 — WER dropped to 7.3%. Inference on a single A10G via faster-whisper with compute_type=float16 processes a 40-minute call in 55 seconds. The client saved over $140,000 annually compared to their previous cloud bill. Contact us for a free pilot estimate to see similar savings on your data.

How to fine-tune Whisper on domain data?

When a general model underperforms, fine-tuning is the first tool. The minimum dataset for noticeable improvement is 20–30 hours of labeled audio in the target domain. Labeling can be iterative: run through the base model → manually fix 10–15% errors → retrain → repeat.

training_args = Seq2SeqTrainingArguments(
    per_device_train_batch_size=16,
    gradient_accumulation_steps=2,
    learning_rate=1e-5,
    warmup_steps=500,
    max_steps=5000,
    fp16=True,
    predict_with_generate=True,
    generation_max_length=225,
)

Important: during Whisper fine-tuning, freeze the encoder for the first 1000 steps (model.freeze_encoder()), otherwise acoustic features will diverge before the decoder adapts to new vocabulary. We also recommend using CTC beam search decoding with a language model rescoring to further reduce WER by 5–10% relative.

Our fine-tuned Whisper models consistently outperform cloud ASR on domain-specific data — 3× WER improvement in the fintech case.

Speech synthesis: How to choose a model for your task?

For interactive bots requiring TTFB < 300 ms — Silero or Kokoro. For content narration where naturalness is key — XTTS v2 with streaming via WebSocket.

Our process and deliverables

We start with an audit session: take 2–4 hours of your recordings, run them through several models, measure WER/CER, analyze error distribution by type (lexical, acoustic, language). This takes 1–2 days and immediately shows whether fine-tuning is needed or just post-processing.

Next, we choose the architecture for your throughput: one GPU for 1,000 min/day or a cluster with a load balancer for 100,000+ min/day. Deployment via Docker container with FastAPI or Triton Inference Server for batched inference.

What you get after engagement:

• Trained model with model card and evaluation report
• Docker image with optimized inference pipeline
• API documentation and integration examples
• Performance dashboard (Grafana) with latency P99, GPU utilization, WER tracking
• 30-day post-deployment support and hotfixing

Timelines depend on complexity:

• Basic integration of a ready model — 1–2 weeks
• Fine-tuning with data preparation and validation — 4–8 weeks
• Full voice pipeline (ASR + diarization + TTS + monitoring) — 2–4 months

Project investments typically range from $20,000 to $80,000. Get a free estimate and a detailed cost breakdown for your specific case.

Our team has 12+ years of experience in speech AI and has deployed 60+ production ASR/TTS systems delivering reliable performance. Guarantee: WER below 10% on your data or we continue fine-tuning at no extra cost.

Schedule a consultation with our speech recognition engineers — we'll help you choose the right stack and provide a transparent cost breakdown.