[SayamDev]

Spoken journey replies repeated the departure time and read ISO dates/ranges as written. This adds shorter contextual replies (pickup, drop-off, wait/get-ready, late/passed times and low confidence) and an optional natural British Kokoro voice. Every time, window and confidence still comes from the existing deterministic engine; neither prediction nor the text form changes.

Design decisions:

• English understanding stays in the existing deterministic action/parser pipeline. Speech templates read the engine directly, including data-limit warnings; no LLM interprets journeys or changes numbers.
• Kokoro generates audio locally in a separate worker after explicit opt-in. Supported GPUs use fp32/WebGPU (~350 MB); other browsers use q8/WASM (~115 MB), plus the speech runtime. Screen results remain immediate. British/enhanced local browser voices provide the default and failure/loading fallback.
• Short chunks prevent tokenizer truncation. Cancellation, stale request ids, bounded load/generation/playback and late voice discovery protect interruptions and lifecycle changes. Cached files recreate the runtime on reopening; readiness is never inferred from a saved preference. The UI identifies the voice actually used and explains playback/timeout fallback.
• No audio or transcripts are stored/logged. Only the voice preference and isolated model/style caches persist. Hugging Face sees public file requests/IP, not journey text. Clear all data also removes model/style files; the service worker preserves these and only removes obsolete shell caches.
• VoiceStudio was considered but requires visitors to run a local API. No hosted voice API, account, key, cloning or new backend is introduced. Original app source remains MIT; the optional combined worker contains GPL eSpeak. Licence notices, source archives and a worker banner accompany it. Node-only dependency overrides resolve sharp/global-agent advisories. CSP now permits self/blob audio and same-origin workers without allowing remote scripts/audio.

Validation: lint, strict TypeScript, production build and npm audit (zero vulnerabilities) pass. The expanded suite has 398 tests covering contextual numbers/dates, midnight windows, warnings, cache/runtime restoration, intentional cancellation, stale audio, device fallback and voice discovery, plus the existing full utterance-to-engine path. An evening run exposed two existing timetable fixtures whose arrivals moved past midnight; they now use a fixed midday test clock, with no production change. Chrome production preview verified model download/warm-up, actual WebGPU natural playback, cached reload, Stop, 390px layout and 48px touch targets with no horizontal overflow.

Limits: CPU speech can miss the deadline and fall back; older phones may have memory/battery constraints and gaps between sentences. Physical iOS/Android autoplay/lockscreen, noisy-car recognition and subjective listening tests remain pending. Free-data/callsign limits remain unchanged. Further speech quality/latency work needs device testing and an efficient on-device model; missing flight/queue/traffic facts need reliable licensed data. Rollback is a PR revert, with no journey-schema migration.