Repository navigation
Conversation
calesthio
left a comment
There was a problem hiding this comment.
Thank you for contributing the Audexum integration. I reviewed the provider, tests and setup guidance against docs/PR_REVIEW_GUIDE.md; the seven focused mocked tests pass. Two contracts need correction before approval.
[P1] Do not report paid synthesis as free. AudexumTTS has no estimate_cost override, so 5,000 characters return 0.0 from BaseTool.estimate_cost, and successful results also default to zero cost. Audexum's API reference documents per-character tier multipliers, and its plans include paid usage. This bypasses budget/approval decisions and understates actual spend. Implement a documented estimate or raise PriceQuoteRequired; report unpriced charges as unknown, rather than zero. Add budget-facing coverage.
[P1] Validate the actual public API contract. The adapter sends xi-api-key to /v1/text-to-speech/{voice_id} with language_code and an output_format query. The current maintained reference instead documents Bearer authentication and POST /api/synthesize with voice, lang and format. Please adapt to that contract, or provide maintained documentation and transport fixtures for the claimed ElevenLabs-compatible route. The supplied mocks only assert the adapter's own assumptions, so they cannot establish compatibility.
No real synthesis or paid API call was made. Please also rebase against current main and run the provider/selector contract suites after these corrections.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Review of OpenMontageInc#836 found two contract problems: - Paid synthesis reported as free. estimate_cost now raises PriceQuoteRequired naming the exact credit count (1 / 1.5 / 2 credits per character on v1 / v2 / v3, per audexum.com/help/credits/how-credits-work); the EUR price of a credit depends on the account's plan or pack, so results carry cost_usd=None, cost_status "unquoted" and estimated_credits, following mai_tts and the eleven_v4 path in elevenlabs_tts. - Undocumented transport. The adapter now calls the documented POST /api/synthesize with Bearer auth and body fields text, voice, lang, format and backend (audexum.com/docs), instead of the ElevenLabs-style route. Also: engine input (v1/v2/v3, inferred from the voice), language region suffixes dropped (en-US -> en), output_format mapped by prefix with unsupported formats rejected before any request, correct file extension, output_schema, docs and .env.example updated, tests for the request contract, budget behaviour (dry_run, scoring) and selector routing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01No5BjeaLzSDP4LhEthEGit
d072b27 to
21f4952
Compare
|
Thanks for the careful review. Both P1s are addressed in 21f4952, on top of a rebase onto current main. Cost. Contract. The adapter now calls the documented Ran: |
# Conflicts: # docs/PROVIDERS.md
|
@calesthio friendly ping: both P1s from your review are addressed (cost is now reported as unquoted rather than 0, and the adapter uses the documented |
Adds
audexum_tts, a TTS provider tool next toelevenlabs_tts/fish_audio_tts.Disclosure: I work on Audexum (https://audexum.com), a TTS/STT API.
What it does
tools/audio/audexum_tts.py, sameBaseToolshape aselevenlabs_tts. It calls Audexum's ElevenLabs-compatible routePOST /v1/text-to-speech/{voice_id}withxi-api-key, usingrequests(already used by the other tools). No new dependency.AUDEXUM_API_KEY; optionalAUDEXUM_BASE_URL.text(max 5000 chars),voice_id(defaultNova),language_code,output_format(mp3_44100_128orwav),output_path. Voice cloning, SSML and timestamps are not supported and the tool does not claim them.estimate_costis not overridden (returns the base 0.0), because I did not want to hard-code a price in the repo. The docs link to the pricing page.tts_selectordiscovers it through the registry; nothing to wire.Free tier: 2,000 credits at signup, then 250/day, no card. Roughly 1 credit per character on standard voices, 2 on V3 voices.
Other files:
.env.example,docs/PROVIDERS.md(section, key table, capability table), provider set intests/contracts/test_phase3_contracts.py, newtests/tools/test_audexum_tts.py(7 tests: key gating, request shape, text limits, HTTP error).Tests:
pytest tests/tools/test_audexum_tts.py tests/tools/test_fish_audio_tts.py tests/contracts/test_phase3_contracts.pygives 113 passed; the 12TestVeoVideofailures also fail on a cleanmainin my environment (missing Google dependency). HTTP is mocked; I have not generated a full video with it.Live check (2026-10-22): the endpoint used here returns audio for every voice listed above.