[bot] HuggingFace Hub: `InferenceClient.automatic_speech_recognition()` not instrumented

[braintrust-bot]

Summary

The huggingface_hub integration instruments four InferenceClient / AsyncInferenceClient task methods — chat_completion, text_generation, feature_extraction, and sentence_similarity — but does not instrument automatic_speech_recognition(). This method generates a text transcript from an audio input and is a stable, production generative execution surface in the HuggingFace Hub SDK.

The OpenAI integration already instruments the equivalent surface (audio.transcriptions.create() via TranscriptionsPatcher), making this an asymmetry across providers.

What is missing

Method	Task	Instrumented?
InferenceClient.chat_completion()	Chat generation	Yes
InferenceClient.text_generation()	Text generation	Yes
InferenceClient.feature_extraction()	Embeddings	Yes
InferenceClient.sentence_similarity()	Semantic similarity	Yes
InferenceClient.automatic_speech_recognition()	Audio → text transcription	No
AsyncInferenceClient.automatic_speech_recognition()	Audio → text transcription (async)	No

automatic_speech_recognition() is distinct from the three tasks tracked in issue #487 (text_to_image, image_to_text, text_to_speech): it takes audio as input and produces a text transcript, which is the reverse modality direction from text_to_speech and involves a different class of execution.

What should be instrumented

Calls to InferenceClient.automatic_speech_recognition(audio, ...) should produce a span capturing:

Span field	Content
input	Audio file path, bytes, or URL
output	text field from SpeechToTextOutput
metadata	provider: "huggingface_hub", model (from model param), task-specific params
metrics	Latency (time_to_first_token), response timing

Both sync and async variants need patchers, consistent with the existing pattern in py/src/braintrust/integrations/huggingface_hub/patchers.py.

Braintrust docs status

not_found — The Braintrust HuggingFace Hub integration docs describe tracing for chat_completion, text_generation, and embeddings. No mention of audio transcription or ASR.

Upstream sources

• huggingface_hub.InferenceClient.automatic_speech_recognition — sync method, available since huggingface-hub ≥ 0.15; current matrix version 1.17.0
• huggingface_hub.inference._generated._async_client.AsyncInferenceClient.automatic_speech_recognition — async variant
• Returns SpeechToTextOutput with text: str
• HuggingFace Hub inference API docs: https://huggingface.co/docs/huggingface_hub/package_reference/inference_client

Relationship to existing issues

• Issue [bot] HuggingFace Hub: InferenceClient generative media tasks (text_to_image, image_to_text, text_to_speech) not instrumented #487 tracks text_to_image, image_to_text, and text_to_speech — a different set of generative media tasks on the same client. This issue covers the audio-to-text transcription surface, which is absent from [bot] HuggingFace Hub: InferenceClient generative media tasks (text_to_image, image_to_text, text_to_speech) not instrumented #487's scope.
• The OpenAI integration instruments audio.transcriptions.create() via TranscriptionsPatcher as the direct analog.

Local files inspected

• py/src/braintrust/integrations/huggingface_hub/patchers.py — defines 8 patchers (4 sync + 4 async) for chat_completion, text_generation, feature_extraction, sentence_similarity; no reference to automatic_speech_recognition
• py/src/braintrust/integrations/huggingface_hub/tracing.py — wrapper implementations for the 4 covered tasks; no ASR wrapper
• py/src/braintrust/integrations/huggingface_hub/integration.py — registers patchers; no ASR patcher
• py/src/braintrust/integrations/huggingface_hub/test_huggingface_hub.py — no ASR test cases
• py/src/braintrust/integrations/openai/patchers.py — TranscriptionsPatcher and AsyncTranscriptionsPatcher instrument audio.transcriptions.create() as the OpenAI analog
• py/pyproject.toml — matrix version huggingface-hub==1.17.0