32 lines
3.6 KiB
Markdown
32 lines
3.6 KiB
Markdown
---
|
|
name: tts-endpoint-parameter-capability-audit
|
|
description: "Determine which request parameters a fleet TTS endpoint actually honors by probing façade and engine separately, with a nondeterminism control, instead of trusting docs or byte diffs."
|
|
---
|
|
|
|
Use when asked whether a TTS (or other OpenAI-compatible shim) endpoint supports a parameter — mood, tone, instructions, speed, seed — or when a request "seems to" change the audio. Docs and single-request comparisons both lie; audit the layers.
|
|
|
|
## Why the naive checks fail
|
|
- **Byte-size diffs are meaningless.** Synthesis is nondeterministic: three byte-identical requests to frank-tts came back 107564 / 134444 / 111404 bytes (~±15 % length). "The mood file differs from the baseline" proves nothing.
|
|
- **HTTP 200 proves nothing.** A shim that rebuilds its own payload answers 200 to `speed:"banana"` because it never looks at the field.
|
|
- **A façade can drop fields the engine supports**, so testing only the published port under-reports capability.
|
|
|
|
## Procedure
|
|
1. **Read the shim's parse block** (which request keys it reads) — not its docstring:
|
|
`ssh -o BatchMode=yes dankclaw 'podman exec frank-tts grep -nE "body\.get|payload = |json\.loads" /app/app.py'`
|
|
Whatever keys appear there (plus what it forwards in its own `payload = {...}` dict) is the real published contract.
|
|
2. **Enumerate the engine's key table from the binary** (works with no `strings` in the image):
|
|
`podman exec <engine> sh -c 'grep -aoE ".{60}instructions.{60}" /app/tts-server | head -2'`
|
|
Adjacent tokens come out as a run — that run IS the accepted-key list (`...instructions response_formatspeedseedmax_new_tokensrepetition_penalty subtalker_top_k subtalker_temperature ...`). Also `podman exec <engine> /app/tts-server --help`, and `podman inspect <c> --format "{{json .Config.Entrypoint}}"` to find the binary.
|
|
3. **Reach the engine directly** when its port is unpublished: `TTS_UPSTREAM` in `podman inspect <shim>` gives the real in-network address (`http://frank-tts-engine:8880`, not `host.containers.internal`). No curl/python on the host? Pipe a script into the container that has one: `podman exec -i frank-tts python3 - <<'PY' ... PY`.
|
|
4. **Control for nondeterminism before claiming an effect**: repeat the identical request twice as the control; only compare conditions separated by far more than the jitter, and prefer an error response (a 400 naming the unsupported flag) as the decisive evidence.
|
|
|
|
## Findings for the dank fleet (verify before trusting — they move)
|
|
- Published façade `:7772` reads only `model`, `input`, `voice`/`voice_id`, `response_format`; it rebuilds the payload, so `speed` and every unknown field are dropped before the engine.
|
|
- Engine (`frank-tts-engine`, qwentts.cpp, qwen-talker-1.7b-**base**, Vulkan) accepts `instructions`, `speed`, `seed`, `max_new_tokens`, `repetition_penalty`, `subtalker_top_k/_temperature/_top_p`, but answers `400 "--instruct is not supported for base models"` for `instructions` on the base checkpoint. `seed` does not make output reproducible.
|
|
- Therefore tone/mood must be produced in the **text layer** (prompt directives), not the TTS request; and `speed` needs a two-line façade fix if tempo control is ever wanted.
|
|
|
|
## Verification habits
|
|
- POST with a wrong *type* (`speed:"banana"`) to learn whether anything validates, and with a bogus `voice` to see the error shape (`available_voices` came from the engine, not the shim).
|
|
- Plain GET of `/v1/audio/voices` for the roster.
|
|
- Read-only probes of the fleet are always allowed; synthesis requests are transient GPU work, keep them to a handful.
|