[bot] Vercel AI SDK `experimental_generateSpeech` and `experimental_transcribe` not instrumented

[braintrust-bot]

Summary

The Vercel AI SDK (ai package) exports experimental_generateSpeech (text-to-speech) and experimental_transcribe (speech-to-text) as core functions. Neither the wrapper (wrapAISDK) nor the auto-instrumentation plugin instruments them. Calls to these functions produce no Braintrust spans.

Every other AI SDK core execution function — generateText, streamText, generateObject, streamObject, embed, embedMany, rerank, and the Agent class methods — is fully instrumented, making these the only uninstrumented generative execution functions in the AI SDK plugin.

What instrumentation is missing

• Auto-instrumentation config (js/src/auto-instrumentations/configs/ai-sdk.ts): No config entries for experimental_generateSpeech or experimental_transcribe. Configs exist for 13 other AI SDK functions.
• Channels (js/src/instrumentation/plugins/ai-sdk-channels.ts): No channel definitions for speech generation or transcription. 14 channels are defined for other functions.
• Plugin (js/src/instrumentation/plugins/ai-sdk-plugin.ts): No handler for speech or transcription calls.
• Vendor types (js/src/vendor-sdk-types/ai-sdk-common.ts): AISDKNamespaceBase does not include experimental_generateSpeech or experimental_transcribe.

A grep for generateSpeech or transcribe across js/src/ returns zero matches.

Upstream reference

Both functions are documented core exports of the ai package (v6+):

Function	Description	Reference
experimental_generateSpeech	Converts text to speech audio using a SpeechModelV3. Accepts model, text, voice, outputFormat, instructions, speed, language. Returns generated audio as base64/Uint8Array with media type.	https://ai-sdk.dev/docs/reference/ai-sdk-core/generate-speech
experimental_transcribe	Generates a text transcript from audio input using a TranscriptionModelV3. Accepts model, audio (string/Uint8Array/ArrayBuffer/URL). Returns transcript text, segments with timing, detected language, and duration.	https://ai-sdk.dev/docs/reference/ai-sdk-core/transcribe

Both are imported from 'ai' (the same package instrumented by this repo) and are available in AI SDK v6, which is already instrumented for other functions (ToolLoopAgent.generate, ToolLoopAgent.stream).

Note: Both carry the experimental_ prefix, indicating the API surface may evolve. However, they are documented, functional, and available in the same SDK version that this repo already instruments for other v6-only features.

Braintrust docs status

not_found — The Braintrust AI SDK integration docs do not enumerate specific supported functions. generateSpeech and transcribe are not mentioned.

• https://www.braintrust.dev/docs/integrations/sdk-integrations/index.md

Precedent in this repo

• AI SDK generateImage is tracked as a separate gap in Vercel AI SDK generateImage not instrumented (wrapper and auto-instrumentation) #1626.
• OpenAI audio APIs (audio.speech.create, audio.transcriptions.create) are tracked in OpenAI audio.transcriptions.create() and audio.speech.create() not instrumented (wrapper and auto-instrumentation) #1629.
• Mistral audio APIs are tracked in [bot] Mistral Audio APIs (audio.speech and audio.transcriptions) not instrumented #1749.
• Groq audio APIs are tracked in Instrument groq audio APIs #1864.
• The AI SDK plugin already instruments 13 other core functions across text generation, structured output, embeddings, reranking, and agent execution — speech/transcription is the remaining uninstrumented modality.

Local files inspected

• js/src/auto-instrumentations/configs/ai-sdk.ts — no speech/transcribe configs
• js/src/instrumentation/plugins/ai-sdk-channels.ts — no speech/transcribe channels
• js/src/instrumentation/plugins/ai-sdk-plugin.ts — no speech/transcribe handlers
• js/src/vendor-sdk-types/ai-sdk-common.ts — no speech/transcribe types
• js/src/wrappers/ai-sdk/ai-sdk.ts — no speech/transcribe wrapping
• e2e/scenarios/ai-sdk-instrumentation/ — no speech/transcribe test scenarios