feat: add voice providers and fix research provider mappings

[dittops]

What

Adds speech-to-text and text-to-speech models to the catalog, and fixes two mapping bugs found while validating the merge against live upstream data.

1. Voice providers

Adds deepgram, elevenlabs, assemblyai, aws_polly and groq to LITELLM_TO_TENSORZERO / TENSORZERO_PROVIDERS, with matching STRIP_PREFIXES entries (all five key their LiteLLM entries as {provider}/{model}, so the prefix is stripped rather than doubled). deepgram, elevenlabs and groq also map into LITELLM_TO_RESEARCH, where truefoundry/models carries cost data for them.

63 voice models now in the catalog:

Provider	Models	Modes
deepgram	42	audio_transcription
groq	10	chat 6, audio_speech 2, audio_transcription 2
elevenlabs	5	audio_transcription 3, audio_speech 2
aws_polly	4	audio_speech
assemblyai	2	audio_transcription

e.g. deepgram/nova-3 → input_cost_per_second: 7.167e-05; elevenlabs/eleven_multilingual_v2 → input_cost_per_character: 0.0001.

2. Fix: azure and xai research mappings

LITELLM_TO_RESEARCH pointed at azure-openai and x-ai; the real truefoundry/models directories are azure-open-ai and xai. The lookup silently never matched, so both providers kept LiteLLM's costs and skipped deprecation filtering entirely.

Fixing it matches 117 azure and 55 xai models.

3. Fix: coerce string costs from research YAML

truefoundry writes small costs as 5e-7. YAML 1.1 requires a decimal point in scientific notation, so PyYAML parses those as strings — and the merge was replacing LiteLLM's numeric cost with a string, breaking any downstream arithmetic.

This already affected 166 values on main (openai, anthropic, gemini, mistral, together_ai). The azure/xai fix above would have pushed it to 259. Now 0.

Why

The voice providers were the ask; the two fixes fell out of verifying it. The string-cost bug is pre-existing on main and independent of the voice work — it just gets worse without the fix, so it ships here.

Reviewer notes

Two output changes that will look like regressions in a naive diff but are intended:

• 41 azure models disappear from the catalog. They're deprecated and were never being filtered, because the typo meant the lookup never matched. Catalog total still rises: 1438 → 1460.
• cost_fields_updated drops 366 → 133. The string-vs-float mismatch was counting '5e-7' != 5e-7 as an update on every merge. The stat now reflects real changes.

Deliberately not included:

• sarvam and tencent — on the voice provider list and present in LiteLLM, but all four entries are mode: chat, not audio. Adding them would put chat models under providers that bud-connect routes to WaaV.
• moonshotai → moonshot-ai is left as-is. Not a typo — upstream has no moonshot directory at all.
• Cartesia, LMNT, Murf, PlayHT, Speechmatics and the rest of the voice list have no entries in either upstream source.

groq is added whole, which brings 15 chat models alongside its 5 audio ones. They're real Groq models with real pricing; filtering a provider to audio-only seemed the stranger choice. Easy to gate if you'd rather.

Downstream

This affects bud-connect's NO_MODEL_PROVIDERS (constants.py). Those five providers now have catalog models, so the seeder's model walk reaches them and the explicit entries become redundant — harmless, but the comment's premise no longer holds for them, and tests/test_voice_providers.py likely needs updating. Separate repo, separate change.

Testing

Verified end-to-end against live LiteLLM and truefoundry/models data, before/after diffed.

• 70 tests pass (9 new), merger.py at 100% coverage
• ruff check and mypy clean

Note: the dev machine only has Python 3.9 and the project requires ≥3.10, so the suite was run against a shimmed copy outside the repo. CI covers 3.10–3.13 properly.

🤖 Generated with Claude Code