Two Distinct Voice Modes

Voice capability in a chatbot isn't a single feature — it's two independent pipelines that can be enabled separately or together depending on your use case.

Voice Input — Whisper Transcription

When voice input is enabled, a microphone button appears in the chat widget. The user taps it, speaks their question, and the audio is transcribed to text using OpenAI Whisper — one of the most accurate speech-to-text models available. The transcribed text is then processed exactly like a typed message.

From the AI agent's perspective, nothing changes: it still receives text and responds with text. The voice input layer is a pure accessibility enhancement at the input stage.

Voice Reply — Text-to-Speech

When voice reply is enabled, the agent's text responses are automatically converted to spoken audio using OpenAI's TTS model. A speaker button appears on each AI message bubble, and responses can be set to auto-play when the user has asked a question via the mic.

The result is a fully conversational loop: speak → AI thinks → AI speaks back.

When to Enable Voice Features

Not every chatbot deployment benefits equally from voice. Here's a practical breakdown of the use cases where each mode adds the most value.

Voice Input is most valuable when:

• Mobile is your primary audience — typing on a small screen is slow. A mic tap is faster.
• Accessibility matters — users with motor impairments, dyslexia, or low literacy benefit significantly from speaking instead of typing.
• Complex queries are common — explaining a problem verbally is often faster and more natural than typing a multi-sentence description.

Voice Reply is most valuable when:

• Hands-free operation is needed — kiosk terminals, reception tablets, in-car interfaces, or smart display devices where the user can't type.
• Hospitality and service environments — a spoken welcome message or spoken booking confirmation feels premium and attentive.
• Accessibility — users with visual impairments can interact entirely without reading the screen.
• Short, high-value responses — TTS works best for concise answers. A 30-second spoken response feels natural; a 3-minute monologue does not.

Choosing a Voice for Your Brand

ChatNexus offers eight distinct TTS voices, each with a different character. Choosing the right one has a real impact on how users perceive your brand.

Voice	Character	Best for
Alloy	Neutral, balanced	General-purpose; safe default for any industry
Ash	Male, clear	Technical support, SaaS, professional services
Ballad	Male, warm	Hospitality, onboarding, welcoming contexts
Coral	Female, friendly	Retail, e-commerce, consumer brands
Echo	Male, confident	Finance, real estate, authoritative contexts
Sage	Female, calm	Healthcare, wellness, legal, careful contexts
Shimmer	Female, expressive	Beauty, lifestyle, creative industries
Verse	Male, versatile	Education, content platforms, media

You can preview each voice from the Voice Settings section of your agent configuration before committing. Try a few with a sample response from your actual knowledge base — the difference becomes obvious quickly.

Speed Controls

TTS responses can be played back at speeds ranging from 0.5× (50%) to 2.0× (200%) of the natural rate, adjustable in 10% increments. The default is 1.0× (100%).

Consider adjusting speed for your use case:

• Slower (0.75×–0.9×) — healthcare or elderly-oriented services where clarity is more important than speed
• Default (1.0×) — appropriate for most contexts
• Faster (1.1×–1.3×) — power users or platforms where users already know the content (e.g., brief notification-style confirmations)

How TTS Costs Work (and Why Caching Matters)

Voice reply adds a TTS API call on top of the standard LLM text generation cost. OpenAI's TTS pricing is based on the number of characters in the text being synthesised.

ChatNexus mitigates this in two ways:

Smart caching

Synthesised audio is cached for 7 days per unique combination of voice, speed, and text. If 100 users ask "What are your opening hours?" and the answer is the same each time, the TTS API is called once — not 100 times. Cached responses are served instantly with no API latency.

Streaming segmentation

For longer responses, audio synthesis begins on the first sentence while the rest of the text is still being generated. This means the user starts hearing the response in under a second rather than waiting for the full text to complete before audio begins.

Voice Input and User Consent

Microphone access requires explicit browser permission. ChatNexus handles this with a built-in consent banner that appears the first time a user taps the mic button. The user must check a consent checkbox before recording begins. Consent is stored locally per agent so returning users aren't prompted again.

Audio recorded for transcription is not stored by ChatNexus. It is sent directly to OpenAI Whisper for transcription and the raw audio is discarded immediately after the text transcript is returned.

Enabling Voice on Your Agent

Voice settings are configured per agent in the Voice Settings section of the agent configuration page. The controls are:

• Voice Input toggle — enables the microphone button in the chat widget. Available on all plans.
• Voice Reply toggle — enables TTS playback for AI responses. Requires Starter plan or above.
• Voice selection — choose from eight distinct voices with real-time preview.
• Speed slider — adjust playback speed from 50% to 200% in 10% steps. Updates live as you move the slider.

Voice doesn't make a chatbot smarter — it makes it feel more human. And in a world where users are comparing your AI assistant to every voice interface they've ever used, feeling human matters.