Spark

Spark voice ask (binary / dump Q&A)

Speak to a local SPARK_BC dump or analysis folder — ears (STT) → question → model → speaking (TTS).

Inspired by clear “ask the binary” product loops elsewhere; not an OpenBin SaaS clone and not gated on OpenBin login. OpenBin remains a separate research/lab note (research/LLM_DECOMPILE.md).

Dump / --compile / --run-bc stay SoT. Companions: ./spark-stt-tts · brain: owned spark-coder (TinyCoder) and/or dump-fact answers · map: Model aspects · voice surface: VOICE.md.

Quick start

make helpers
        make spark-stt-tts # optional for live STT/TTS
        # optional tiny brain:
        ./spark-code train --scale tiny

        # Text (no mic) — dump facts / TinyCoder
        ./spark-ask docs/examples/spark-train-step.sparkbc \
         --text "What opcodes are in this dump?"

        # Voice loop, offline CI (stub STT + stub WAV)
        ./spark-speak-ask docs/examples/spark-train-step.sparkbc --dry

        # Voice loop, live mic + local synth (needs spark-stt-tts)
        ./spark-speak-ask docs/examples/spark-train-step.sparkbc --mic

        # Analysis folder from project loop (sibling spark-analyze)
        ./spark-ask out/analyze/demo --text "How many opcodes?"
        ./helpers/spark-speak-ask out/analyze/demo --dry

Root ./spark-ask and ./spark-speak-ask are thin aliases of helpers/spark-ask / helpers/spark-speak-ask.

Modes

Flag Behavior
--text Q Skip mic; answer Q
--voice / spark-speak-ask STT → answer → TTS
--dry Fixture question + stub WAV (CI / no sidecar)
--wav FILE Listen from RIFF/WAVE
--mic arecord via ./spark-stt-tts
--out-wav PATH Write spoken answer
--play aplay after live speak
--weights PATH TinyCoder safetensors
--json Machine-readable result

Answer quality

  1. Dump facts — opcodes, sha256, magic SPBC, TRAIN presence answered from decoded dump (preferred SoT).
  2. TinyCoder — when models/spark-coder/weights.safetensors exists, inject dump excerpt + question; greedy generate.
  3. Weak / missing — note: TinyCoder is tiny, not OpenBin-level RE Q&A, Prefer dump.txt / ops.json.

STT / TTS gates

Same gates as VOICE.md:

Gate Meaning
(default local) spark-stt-tts PCM synth; whisper / SPARK_STT_CMD when available
SPARK_STT_NET=1 + URL HTTP STT
SPARK_TTS_NET=1 + URL HTTP TTS
SPARK_SPEECH_NET=1 Both

URL without gate → fail closed (exit 2). Never print Bearer keys.

Tests

make test-spark-ask
        # or:
        PYTHONPATH=python:tools python3 -m unittest \
         tools.spark_ask.test_voice_ask -v

Grounding / anti-guess

Dump facts are the SoT for opcode / sha answers. For inventable prices or free generate, use forced grounding — abstain unless expect / fixture / dump match:

./spark-ground ask --prompt "…" --dump dump.txt \
         --candidate "…" # miss → exit 2
        make test-ground

See knowledge/SAFETY_LIMITS.md (Grounded generation / anti-guess). Recompile ≠ semantics.