Share
Embed
Chatless-AI - Realtime-Sprachinterfaces für Web und Mobile entwickeln
Mit Realtime-Sprachmodellen wie GPT-realtime oder Gemini-Live entsteht eine neue Generation von Interfaces: Sprache wird zum sofort reagierenden, latenzarmen Interaktionskanal – ohne Prompting, ohne Wartezeiten, hands-free.
In diesem Talk zeigt Sascha Lehmann, wie Realtime-Modelle technisch funktionieren, wie man Kontextgrenzen, Rollen und Sicherheit zuverlässig kontrolliert und wie sich Realtime-AI gezielt in Web- und Mobile-Anwendungen integrieren lässt – von Architektur über Kostenoptimierung bis hin zur UX, die Nutzer transparent durch den Dialog führt.
Sascha Lehmann
October 01, 2026
Transcript
35 slides
-
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Sascha Lehmann @derLehmann_S Consultant -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Sascha Lehmann Consultant @ Thinktecture AG @derLehmann_S https://www.linkedin.com/in/sascha-lehmann [email protected] https://www.thinktecture.com/thinktects/sascha-lehmann/ -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile What ”real-time” really means Speech interaction until now • Each arrow represents 200-500ms of latency • Like a zoom call with huge delay Mic STT (Speech to Text) LLM TTS (Text to Speech) Speaker -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile What ”real-time” really means With realtime model • Permanent available audio channel • It no longer feels like software anymore, but like a real counterpart • Average response time ~320ms (OpenAI figure for gpt-realtime); full-duplex models like gpt-live-1 listen while they speak Mic STT (Speech to Text) Realtime Model TTS (Text to Speech) Speaker -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Available Realtime Models Property OpenAI gpt-realtime-2.1 OpenAI gpt-realtime-2.1-mini OpenAI gpt-live-1 Gemini 3.8 Live Architecture Speech-to-Speech native (speech, reasoning, tools in one model) Speech-to-Speech native Full-duplex voice layer, reasoning & tools delegated to backend model Multimodal Live (Audio/Video/Text) Latency ~320ms first response (per OpenAI) ~320ms Full-duplex: listens while speaking Sub-800ms Languages Multilingual, language change midsentence Multilingual Multilingual, launch voices skew English 70 languages Particular strengths GPT-5-class reasoning, best instruction adherence, tool calling Cost efficient, quick, reasoning included Natural interruptions, noise handling, free choice of backend model Video input, multimodal, cost efficient Barge-In support (native, full-duplex) Emotional adaption Configurable via prompt restricted Tone, pace & style via prompt Affective Dialog native Voices 10+ voices Like gpt-realtime 12 voices (new accents & dialects) 30+ HD Voices Latest models gpt-realtime-2.1 (Jul 2026); gpt-realtime retires Jan 20, 2027 gpt-realtime-2.1-mini (Jul 2026) gpt-live-1 (Sep 10, 2026, GA) gemini-3.8-live (stable since Sep 15, 2026) -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Architecture High Level • Model processes audio directly (lowest latency) • Thinks and responds in speech • Doesn’t rely on transcript • Hears emotion and intent • Filters noise • Two paradigms since 2026: gptrealtime-2.x = one model for speech, reasoning & tools. gpt-live-1 = fullduplex voice layer, reasoning & tools delegated to a backend model https://developers.openai.com/api/docs/guides/voice-agents/ -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Architecture Transport – WebRTC vs. WebSocket Aspects WebRTC WebSocket Protocol UDP (packet loss tolerant) TCP (guaranteed delivery) Latency Very low (~50–150ms) Higher (TCP Head-of-Line Blocking) Features Echo Cancellation, Noise Suppression, Jitter Buffer - Ideal for Browser-Apps, Client-side Server-side, Phone-integration Downsides NAT Traversal complexity Latency peaks during packet loss Also available gpt-realtime-2.x, gpt-live-1 OpenAI realtime/live, Gemini Live; SIP for telephony https://developers.openai.com/api/docs/guides/voice-agents/ -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Realtime Connection RealtimeAPI-Key Backend System Prompt 1 Ephemeral Key 2 3 4 App RealtimeAPI WebRTC -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Client side connection -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Sending events -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile UX Challenges -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile 1. What is my current state? • The user has no visual clue • Voice needs a state indicator • Other methods • Audio cues • Sound effects -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile 2. Success / Failure feedback • Unique Audio cue sound • Vocal feedback from the agent • Visual indicator update -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile 3. Error Handling • Automatic retry • Agent asks for clarification • Control to restart the voice agent session -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile 4. Interruption (Barge-In) • User needs the possibility to cancel an action • Voice activity detection (VAD) • Also include manual cancellation -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Realtime Models do not only support speech • Voice is not the only interface – it is an additional one • Realtime models are also ”Multimodal Models” • Speech • Text • Images (gpt-realtime-2.x, Gemini Live) • Gemini Live also supports Live Camera Input - So the model can see what you currently see • Exception: gpt-live-1 is audio-only – no image input; visual context has to go through the backend model -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile GPT-Live Full Duplex: listening while speaking • Turn-based: model waits for silence, then answers • Full-duplex: one model hears and speaks at the same time • User silence → VAD interrupt cut off, wait Agent Handles interruptions and background noise natively, no VAD tuning • Turn-taking latency 0.8 s vs. 1.4 s on gpt-realtime-2.1 • Full Duplex Bench interactivity 80 % vs. 45 % • Turn-based (gpt-realtime-2.x) Same tech behind ChatGPT Voice since Jul 2026, API since Sep 10, 2026 time Full-duplex (gpt-live-1) interjects while agent talks User Agent no gap listens while speaking, adapts, continues time Turn-taking latency and Full Duplex Bench figures: OpenAI, Sep 2026 (not independently reproduced) -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile GPT-Live Architecture: voice layer + delegated backend gpt-live-1 User audio in / out Conversation layer Full-duplex speech, turn-taking, tone & pace, noise handling $0.05 / min delegate result Backend Your agent or a text model GPT-6 Astra, GPT-5.6 or any harness: reasoning, tools, RAG, guardrails billed separately (tokens) Tools APIs, search vs. gpt-realtime-2.x Trade-offs Worth it when • Realtime: one model speaks, reasons and calls tools • Live: talking and thinking are split; it keeps talking while the backend works • Per-minute price instead of audio tokens + Most natural turn-taking; backend is freely swappable – No image/video input, no structured outputs on the voice layer – Two systems to design, monitor and secure • Phone and support agents with frequent interruptions • Long sessions: tutoring, reservations, scheduling • You already own an agent backend (tools, RAG, guardrails) Source: OpenAI, Introducing GPT-Live-1 in the API, Sep 10, 2026 -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile How much does it cost? -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Pricing model of each model • ~10 Token/Sek (Input) • ~20 Token/Sek (Output) • • Cost increase with conversation length Each turn resends the context Modell Input (Audio) Cached Input Output (Audio) ~cost per minute (conversation) gpt-realtime-2.1 $32 / 1M Tokens $0.40 / 1M Tokens $64 / 1M Tokens (Text: $24) ~$0.06–0.12 gpt-realtime-2.1-mini $10 / 1M Tokens $0.30 / 1M Tokens $20 / 1M Tokens ~$0.02–0.04 gpt-live-1 $0.05 / Min (VoiceLayer) – Backend-Modell separat ~$0.05 + Backend Gemini 3.8 Live $3 / 1M Tokens Kein Caching $12 / 1M Tokens ~$0.01–0.02 -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile How to keep the costs under control 1. Make use of prompt caching 2. Make agent not too talkative 3. Truncation & Retention Ratio konfigurieren 4. Manage Conversation history manually 5. Make use of mini model 6. Proactive Tool calling without confirmation 7. Design clear conversation flows with a final end -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Example calculation • Average duration of a pocket depth measurement: ~ 3 min • With gpt-realtime-2.1-mini + caching: ~$0.04–0.08 per finding • With Gemini 3.8 Live: ~$0.03–0.06 per finding -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Integration & Best Practices -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile System Prompt -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile System Prompt Best Practices • Role & Objective — Clear identity + task definition ("You are a dentist assistant, your task is form-filling") • Instructions / Rules — Bullet points over prose, CAPS for critical rules ("NEVER GUESS", "ALWAYS use get_selected_tooth first") • Personality & Tone — Language, formality, response length, pacing explicitly set (e.g. "max 2 sentences", formal address, "professional & concise") • Conversation Flow — State machine for procedures (Pocket Depth: Init → Measurement Loop → Completion) • Reference Pronunciations — Phonetic guides for domain terms + tooth numbers that TTS would mispronounce ("one-eight" not "eighteen") • Tools / Function Calling — Per tool: trigger condition, preamble ("One moment..."), error handling (1x silent retry, 2x notify user, 3x fall back to manual) • Safety & Escalation — Domain constraint, exact refusal scripts, escalation after 3x off-topic • Examples — Concrete input/output pairs for ambiguous cases -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile System Prompt Key Techniques • Bullet Points > Prose — Model follows short bullets better than paragraphs • Labeled Sections — Model navigates by heading; each section = one concern • Lock the Language — "Respond ONLY in German" stated explicitly, prevents drift • Unclear Audio Handling (3-Step) — 1. Ask to repeat → 2. Address audio quality → 3. Offer to skip • Rotating Confirmation Phrases — Provide a list of alternatives ("Understood", "Noted", "Alright") → prevents robotic repetition • Tool Preambles — Short sentence BEFORE the tool call so the user knows what's happening • Confirm Critical Data — For medical values always read back: "Tooth one-eight: three. Correct?" • Strict Domain Constraint — Exact scripted response for out-of-scope requests + escalation threshold -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Security & Guardrails • No API-KEY in CLIENT! → Generate Ephemeral Token via Server • Clear definitions of what is allowed and what is not – System Prompt • Escalation-Tool: If there is uncertainty → Abort or handoff to a human agent • With gpt-live-1 the backend model is yours: guardrails, data access and logging live in your own agent, not in the voice layer • DSGVO: Check in which country the model is running (EU data residency via Azure OpenAI or Vertex AI regions) -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Evaluation How do I know a prompt change actually made things better? -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Evaluation Why it is hard • Two axes, not one — A response can be right and still sound broken. Grade both separately. • Content quality — Did it do the right thing? Correctness, tool choice, tool arguments, instruction following. • Audio quality — Did it sound acceptable? Naturalness, pacing, pronunciation, behaviour under noise. • A turn is a pipeline — speech start/stop, commit, response, audio deltas. Log every stage to find the real root cause. • Transcript is not ground truth — The audio signal is the truth; a transcript is just a model’s interpretation and can be wrong. • False fails & false passes — ASR drops a digit the model heard correctly, or a clean transcript hides clipped audio the model guessed at. • Grade on transcripts + traces — Run most automated grading on transcripts at scale; calibrate graders on noisy, production-like text. • Add an audio audit loop — Spot-check ~1–5% of sessions end-to-end by actually listening. -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Evaluation Crawl / Walk / Run • Build complexity in steps — If your system cannot crawl, it will not run. Early evals must be diagnosable, repeatable and cheap. • Crawl — Synthetic (TTS) audio + single-turn. Tests intelligence: intent routing, tool choice, valid arguments. • Walk — Real noisy audio + single-turn. Tests perception: does it still hear "7pm", not "7" or "7:15"? • Run — Simulated user + multi-turn. Tests robustness: holds the goal, sequences tools, recovers from errors. • Example: "Change my reservation to 7pm" — Crawl grades the next turn only; Walk replays it under phone-bandwidth noise; Run adds messy follow-ups + an injected tool error. • Single-turn vs multi-turn — Single-turn = can you win the battle. Multi-turn = can you win the war (episode-level outcome). -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Evaluation The loop: does a change actually help? • Dataset — Start with a gold seed set (10–50) of must-not-fail flows. Balance positives and negatives; tag by intent, audio, language, expected tool. • The iteration loop — Run evals, localise the failure to one behaviour, change ONE thing, rerun, confirm the fix improved it without regressions. • Graders — Layer them: deterministic (tool calls, JSON, patterns), LLM rubric (correctness, helpfulness), audio (silence, overlap, interruptions). • Regression suite — Hard cases you already fixed. Run on every prompt, model or tool change. Your "do not break" contract. • Harness — One job: make runs comparable. Pin audio bytes, chunking and VAD; prefer VAD off + manual commit for reproducibility. • Manual review is highest-leverage — Automation shows what you can measure; listening shows what you should measure. • Rolling discovery set — Fresh failures from production. Promote real failure modes into the offline dataset over time. • Holdout set — Untouched subset run occasionally. If test scores climb while holdout stays flat, you are training for the test. -
Chatless AI Architecting Realtime, Hands-Free AI Interfaces for Web &
Mobile Summary • Realtime Voice ≠ Chatbot with microphone: It is a complete new interaction paradigm • Speech-to-Speech is the way for low-latency hands-free scenarios • UX is the real challenge – clear and precise communication of state to the human user is key • Cost is controllable – with caching, mini-models, truncation and short sessions • Start small – one scenario, one tool, one demo • And MOST IMPORTANT: The implementation needs to provide a REAL benefit for the user