Big push driven by user feedback ("делай имбу") and web research on what
voice assistants need to be the ideal:
TTS backend abstraction (P0.1)
- new module crates/jarvis-core/src/tts/{mod,sapi,piper,silero}.rs
- TtsBackend trait with SapiBackend (current PowerShell), PiperBackend
(rhasspy/piper, neural quality), SileroBackend (python subprocess)
- JARVIS_TTS env var picks (sapi|piper|silero). Auto-detect Piper if
binary + voice present in tools/piper/. Falls back to SAPI on missing.
- SpeakOpts {lang, detached, raw} replaces ad-hoc args. text_utils
sanitiser applied unless raw=true.
- llm_fallback + lua/api/tts both routed through tts::backend().
- tools/piper/install.ps1 downloads piper.exe + ru_RU-irina-medium.onnx
from rhasspy releases + huggingface. Smoke-test included.
- tools/silero/silero_tts.py helper (PyTorch); rust spawns it as subprocess.
IMBA-1 Agentic LLM router
- crates/jarvis-app/src/llm_router.rs
- When fuzzy/intent matcher fails, LLM picks the closest command from the
full registry. Returns JSON {command_id, confidence, reason}.
- Threshold-gated re-dispatch via substitute phrase. JARVIS_LLM_ROUTER=1
enables; JARVIS_LLM_ROUTER_THRESHOLD overrides 0.55 default.
- Inserted in app.rs::execute_command between "no match" and existing
llm_fallback chat fallback.
IMBA-2 Long-term memory
- crates/jarvis-core/src/long_term_memory.rs — JSON store at
APP_CONFIG_DIR/long_term_memory.json. Atomic write-through.
- remember/recall/search/forget/all/build_context API.
- Lua bindings: jarvis.memory.* (5 functions).
- llm_fallback auto-injects relevant facts (substring search of prompt)
into system message before LLM call.
- Pack resources/commands/memory_pack/ with 4 commands: remember, recall,
forget, list.
IMBA-3 Profile switching (work/game/sleep/driving/default)
- crates/jarvis-core/src/profiles.rs — JSON profiles at APP_CONFIG_DIR/profiles/
Auto-seeds 5 defaults on first run with personality + allow/deny lists +
greetings + emoji icons.
- active_profile.txt persists choice across restart.
- Lua bindings: jarvis.profile.{active,set,list,allows,active_name}.
- llm_fallback prepends profile personality to system prompt.
- Pack resources/commands/profile_switch/ with 6 voice triggers.
IMBA-4 Multimodal screenshot + vision LLM
- crates/jarvis-core/src/lua/api/vision.rs — gated on HTTP sandbox.
- jarvis.vision.screenshot() captures via PowerShell System.Drawing.
- jarvis.vision.describe(prompt?) sends base64 PNG to Groq vision model
(default llama-3.2-11b-vision-preview, override via GROQ_VISION_MODEL).
- Pack resources/commands/vision/ with 2 commands: describe + read_error.
P0.2 Continuous conversation grace window
- config::CONVERSATION_GRACE_MS = 30_000.
- app.rs: after command result, if grace_ms > 0 keep listening WITHOUT
re-wake for the grace duration. Existing CMS_WAIT_DELAY back-dated so
the existing timeout fires at start + grace_ms.
Tests: 24/24 jarvis-core unit tests pass (including 5 text_utils).
Build: cargo build --release -p jarvis-app and -p jarvis-gui both succeed
on Windows MSVC (VS 2026 Enterprise vcvars64).
Notes for setup:
- Piper voice install: pwsh tools/piper/install.ps1 (downloads ~90 MB).
- GROQ_TOKEN needed for IMBA-1 (router) and IMBA-4 (vision).
- All features are opt-in via env vars or auto-detect; existing SAPI +
fuzzy match path remains the default.
User complaint: SAPI reads every dotted acronym character-by-character with a
"точка" between letters. Unbearable on every speak. No real TTS analog
shipping yet (Silero/Coqui/Inworld in roadmap), so the immediate fix is text
preprocessing.
New module `jarvis-core::text_utils` with `sanitize_for_speech()`:
- Specific brand mapping: J.A.R.V.I.S. → Джарвис, U.S.A. → США, U.K. →
Британия, U.S. → США, S.O.S. → сос.
- Generic dotted-acronym collapse: any `[letter].[letter].[letter].(...)`
run of 3+ letters gets its dots stripped. "T.O.N." → "TON".
- URL stripping: anything starting with http/https gets replaced with the
word "ссылка". SAPI reading URLs char-by-char is unlistenable.
- Em/en-dash normalisation: — → -, « » " stripped.
- Whitespace collapse.
Wired in two places:
- `jarvis.speak(text, opts?)` Lua API runs sanitiser unless opts.raw=true.
- `jarvis-app/llm_fallback::speak_via_sapi` runs sanitiser before the SAPI
PowerShell shell-out for LLM auto-fallback replies.
5 unit tests in `text_utils::tests` covering jarvis-name collapse, generic
acronym pattern, URL stripping, dash normalisation, and "leave clean
russian sentences alone" — all green.
Real TTS upgrade (Silero v4 first cut) tracked as P0.1 in roadmap.
GUI front-end does not subscribe to IpcEvent::LlmReply, so the LLM
auto-fallback was effectively silent — user got the OK chime but no
spoken answer. Backlog item resolved.
llm_fallback::handle now also invokes speak_via_sapi() after firing
IpcEvent::LlmReply (both for successful replies and the fallback
error message). The text goes through System.Speech.Synthesis.
SpeechSynthesizer, with a preference pass that picks the first
installed ru-RU voice if present (Microsoft Irina / Pavel etc.) so
Russian replies sound right out of the box, and falls back to the
default English voice otherwise.
PowerShell is invoked fire-and-forget (no Wait, stdout/stderr null)
so the voice loop never blocks. A long answer keeps speaking while
jarvis returns to wake-word listening; the mic may pick up some of
the speech which is the same compromise voices::play_* already makes.
A cleaner version would pause pv_recorder for the duration of the
utterance — that goes in the backlog as part of the eventual
unified IpcEvent::Speak refactor.
Toggle: set JARVIS_LLM_TTS=false to disable (e.g. on a server where
the GUI is the speaker). Non-Windows builds get a no-op stub.
new module crates/jarvis-app/src/llm_fallback.rs holds an optional
LlmClient + ConversationHistory built once at startup. if GROQ_TOKEN is
unset the module logs a warning and stays disabled — voice commands keep
working as before.
both the wake-word voice path (recognize_command in app.rs) and the
gui-side text command path (process_text_command) now check whether the
recognized phrase starts with one of the configured trigger phrases
(ru: 'скажи' / 'ответь' / 'произнеси'). when it does, the remainder of
the phrase is sent to Groq and the reply is published as a new
IpcEvent::LlmReply { text } so the gui can speak it.
on api error the trailing user turn is popped from history, the russian
fallback line is sent over ipc and a 'error' voice cue plays. the loop
itself never panics.
