Voice easy — train STT / TTS in 3 steps

Piece-of-cake path for owned Spark voice-related heads we write and train in this repo. Companion to VOICE.md and Model aspects (ears/speaking).

Owned Spark STT/TTS heads trained in this repo — not a vendor TTS SaaS. Prefer CPU or a consumer GPU for optional train.

3 steps

# 1) Env check (no secrets)
        ./spark-voice env --dry

        # 2) Tiny happy path (CI / laptop) — fixtures → train → prove
        ./spark-voice easy --dry --device auto

        # 3) Status
        ./spark-voice status

Outputs:

models/spark-voice-easy/weights.safetensors (+ arch.json, checkpoint.json)
out/voice_easy/fixtures/ — owned tone WAVs + phrases
out/voice_easy/roundtrip/ — TTS dry WAVs + roundtrip.json

Scales

Scale	How	Dims / steps	Device
tiny (default)	`--scale tiny` or omit	dim 16, ~12 steps	CPU fine; CI
large (opt-in)	`--scale large` or `VOICE_SCALE=large`	dim 256, ~80 steps	Consumer GPU suggested; ~2 GiB VRAM hint

# Large — opt-in; prefers a consumer GPU; fail closed without one
        ./spark-voice easy --scale large --device auto

        # Large on CPU only when you explicitly ask (still owned weights)
        ./spark-voice easy --scale large --device cpu

Large ≠ production vendor quality. It is a bigger owned head for local experiments — covers fixture-scale STT classify + PCM tone TTS.

Fail closed (device guard)

If --scale large and the visible GPU is reserved for other workloads, Spark refuses (exit 2) unless you pass --device cpu. Training never routes to reserved devices.

Flags

Flag / env	Meaning
`--dry`	CI-friendly (clamped steps)
`--device auto\\|cpu\\|…`	device pick; auto prefers a consumer GPU (or CPU)
`--scale tiny\\|large`	model size
`VOICE_SCALE=large`	same as `--scale large` when flag omitted

Reuse ears / speaking

Language listen / speak and ./spark-stt-tts stay the live sidecar path (VOICE.md). Voice-easy trains a separate owned head under models/spark-voice-easy/ so the training path is obvious:

ears (fixtures / spark-stt-tts)
         → voice-easy STT head (classify phrases)
        brain (optional spark-coder / dump ask)
         → voice-easy TTS head (PCM params)
        speaking (roundtrip WAV / spark-stt-tts speak)

External STT / TTS sidecars (later)

Plug vendors outside these owned weights — same gates as voice:

Gate	Meaning
`SPARK_STT_CMD` / `SPARK_TTS_CMD`	Local shell (`%i` / `%o`)
`SPARK_STT_NET=1` + URL	HTTP STT
`SPARK_TTS_NET=1` + URL	HTTP TTS

URL without gate → fail closed. Never commit Bearer keys.

IDE

Spark IDE extension command “Spark: Voice easy train” runs ./spark-voice easy --dry (CLI-first; stub task hook).

CLI / CI

./spark-voice easy --dry # dry tiny happy path
        ./spark-voice easy --scale large --device auto # opt-in local large

Scope

Claim	Status
Owned STT/TTS heads we train	yes (tiny + large)
CI dry green	yes (`make test-voice-easy`)
Vendor mega-TTS overnight	Out of scope
Marketing win banners	Out of scope