feat: add speech clarity audio scorer

[kev (kevchoi)]

AI-created or modified and not human-reviewed in its current form; treat this artifact as provisional. Updated 2026-10-07.

Adds SpeechClarity, an LLM judge that listens to a voice call recording and rates how easily the agent's speech can be understood: garbled or cut-off words, distortion, dropouts, echo and noise. Scores are A = 1, B = 0.5 and C = 0. This PR carries over only the speech clarity part of #228. TurnTaking, VoiceTaskSuccess (#241) and the thread changes stay out.

• templates/speech_clarity.yaml: the prompt, with model: gemini-3.8-flash and a new audio: input.audio field.
• Templates take an optional audio path (JS LLMClassifierFromTemplate, Python LLMClassifier). The value there must be {data: bytes, content_type: "audio/..."} or a list of them, such as a recording split into chunks. Each is sent unchanged as its own base64 file part after the prompt text, in list order; nothing is joined. If the value is missing or the list is empty, the score is skipped (null). Chunk boundaries aren't marked in the prompt, so a word split across two chunks may be judged as cut off.
• js/manifest.ts: lists SpeechClarity as a built-in under "LLM-as-a-Judge". Built-ins get audio as an attachment reference, so it needs the API server to pass the bytes (braintrustdata/braintrust#21432, in the matching braintrust stack). Merge the stacks together.
• SCORERS.md documents SpeechClarity and the audio: field; README.md and AGENTS.md list it.

Only Gemini reads audio this way. Tests mock the API, so the scorer hasn't been run against a real model or calibrated against human ratings.

Testing

• pnpm run test -- js/llm.test.ts: passes.
• uv run --extra dev --extra scipy pytest py/autoevals/test_llm.py -k speech: passes. Five unrelated tests in that file need an OpenAI key and fail without one.
• tsc --noEmit reports one error in js/render-messages.test.ts. This PR doesn't touch that file.

🤖 Generated with Claude Code

[github-actions]

Braintrust eval report

Autoevals (HEAD-1791394521)

Score	Average	Improvements	Regressions
NumericDiff	78.4% (0pp)	9 🟢	10 🔴
Time_to_first_token	8.77tok (+2.17tok)	15 🟢	204 🔴
Llm_calls	1.55 (+0)	-	-
Tool_calls	0 (+0)	-	-
Errors	0 (+0)	-	-
Llm_errors	0 (+0)	-	-
Tool_errors	0 (+0)	-	-
Prompt_tokens	516.87tok (+0.37tok)	22 🟢	23 🔴
Prompt_cached_tokens	0tok (+0tok)	-	-
Prompt_cache_creation_tokens	0tok (+0tok)	-	-
Prompt_cache_creation_5m_tokens	0tok (+0tok)	-	-
Prompt_cache_creation_1h_tokens	0tok (+0tok)	-	-
Completion_tokens	460.81tok (-12.29tok)	112 🟢	99 🔴
Completion_reasoning_tokens	345.89tok (-13.38tok)	89 🟢	77 🔴
Completion_accepted_prediction_tokens	0tok (+0tok)	-	-
Completion_rejected_prediction_tokens	0tok (+0tok)	-	-
Completion_audio_tokens	0tok (+0tok)	-	-
Total_tokens	977.68tok (-11.92tok)	110 🟢	101 🔴
Estimated_cost	0$ (0$)	72 🟢	60 🔴
Duration	8.78s (+2.17s)	15 🟢	204 🔴
Llm_duration	9.63s (+2.52s)	13 🟢	206 🔴