[bot] Instrument aws-sdk-transcribestreaming (4,014 weekly downloads)

[braintrust-bot]

Summary

The Braintrust Rust SDK has no instrumentation for aws-sdk-transcribestreaming, AWS's official Rust SDK crate for the Amazon Transcribe Streaming service — real-time speech-to-text execution. Braintrust already treats audio transcription as an in-scope, traced generative-AI capability (its OpenAI integration docs list "Transcription" as a covered Audio API, and issue #84 in this repo already flags the equivalent gap for local whisper-rs execution), but no Braintrust SDK instruments AWS's hosted real-time transcription service.

What is missing

aws_sdk_transcribestreaming::client::Client exposes these execution-oriented operations (per https://docs.rs/aws-sdk-transcribestreaming/latest/aws_sdk_transcribestreaming/client/struct.Client.html):

• start_stream_transcription — the core execution call: opens a bidirectional stream, accepting an input audio-chunk event stream and returning a live transcript-result event stream, with configurable language identification, vocabulary filters, speaker partitioning, and PII redaction.
• start_call_analytics_stream_transcription — real-time Call Analytics transcription (adds sentiment/categorization signals to the live transcript stream).
• start_medical_stream_transcription — specialty medical real-time transcription.
• start_medical_scribe_stream / get_medical_scribe_stream — Amazon Web Services HealthScribe session execution and status retrieval.

None of this is wired into the SDK:

1. No usage extraction path — extract_openai_usage()/extract_anthropic_usage() in src/extractors.rs parse a JSON usage object with token counts. Transcribe Streaming has no token-usage concept at all (billing is by audio duration, not tokens) and returns typed Rust event-stream structs (TranscriptEvent, Result, Alternative), not a serde_json::Value — neither extractor applies.
2. No stream aggregation path — BraintrustStream::aggregate() (src/stream.rs) deserializes items as StreamChunk { model, choices, usage } (OpenAI Chat Completions shape) from a Stream<Item = Result<Value, E>>. Transcribe Streaming's bidirectional event stream (interim + final Results with per-word timing/confidence) has no choices/JSON structure and is structurally incompatible with the existing aggregator, mirroring the same class of gap already noted for whisper-rs ([bot] Instrument whisper-rs (58,308 weekly downloads) #84) and Gemini's streamGenerateContent ([bot] BraintrustStream cannot aggregate Google Gemini streamGenerateContent events #60), but for AWS's distinct, typed, bidirectional-streaming client.
3. No transcription span type — there is no equivalent of a "transcription span" (input audio stream reference, output transcript segments with timestamps/confidence, detected language) anywhere in src/types.rs.

Library significance

• Official AWS-maintained crate, part of the aws-sdk-rust monorepo (https://github.com/awslabs/aws-sdk-rust), published by the "AWS SDK Rust Bot"
• Actively released: latest version 1.112.0
• API docs: https://docs.rs/aws-sdk-transcribestreaming/latest/aws_sdk_transcribestreaming/
• Crate: https://crates.io/crates/aws-sdk-transcribestreaming

Weekly downloads

Weekly downloads: 4,014 (as of 2026-08-12; https://crates.io/api/v1/crates/aws-sdk-transcribestreaming/downloads)

Methodology: exact sum of the crates.io downloads API's daily version_downloads entries (summed across all versions publishing downloads on a given day) plus meta.extra_downloads entries, for the complete 7-day window 2026-08-06 through 2026-08-12 inclusive (checked on 2026-08-17; more recent days were excluded due to crates.io's daily-aggregation lag). Per-day combined totals: 08-06: 630; 08-07: 527; 08-08: 481; 08-09: 394; 08-10: 679; 08-11: 492; 08-12: 811. Sum = 4,014. 90-day recent_downloads reported directly by crates.io: 51,181 (consistent with this weekly figure, ≈3.9K/week average).

Braintrust docs status

supported for the underlying capability class (audio transcription), but not_found for Amazon Transcribe/AWS specifically. Checked https://www.braintrust.dev/docs/integrations/ai-providers/openai directly on 2026-08-17: the Python span-coverage table lists "Audio APIs (audio captured as an attachment)" as covering "Transcription, Translation, Speech," confirming transcription is already an in-scope traced capability class for Braintrust (for OpenAI's client). Neither Amazon Transcribe nor any AWS transcription service is mentioned on https://www.braintrust.dev/docs/integrations/ai-providers or https://www.braintrust.dev/docs/integrations/ai-providers/bedrock.

Relationship to existing issues

• Distinct from [bot] Instrument whisper-rs (58,308 weekly downloads) #84 (whisper-rs): [bot] Instrument whisper-rs (58,308 weekly downloads) #84 covers fully local/offline transcription via whisper.cpp bindings with no network calls involved. This issue covers AWS's hosted, real-time bidirectional-streaming transcription service with its own distinct client, event-stream protocol, and typed Rust API.
• Distinct from [bot] OpenAI Chat Completions audio output transcript silently dropped from span output #72/[bot] ChatMessage and StreamDelta drop OpenAI audio output (gpt-4o-audio models) #61 (OpenAI Chat Completions audio output): those cover audio generated as part of OpenAI chat/completions responses, not a dedicated transcription execution surface.
• Searched this repository's open issues for "transcribe", "transcription", "start_stream_transcription" — the only match is [bot] Instrument whisper-rs (58,308 weekly downloads) #84 (whisper-rs), which is explicitly local/offline and does not cover this AWS crate.

Upstream sources

• aws-sdk-transcribestreaming on crates.io: https://crates.io/crates/aws-sdk-transcribestreaming
• API docs (Client — start_stream_transcription, start_call_analytics_stream_transcription, start_medical_stream_transcription): https://docs.rs/aws-sdk-transcribestreaming/latest/aws_sdk_transcribestreaming/client/struct.Client.html
• aws-sdk-rust GitHub monorepo: https://github.com/awslabs/aws-sdk-rust
• crates.io downloads API (raw data used for the weekly-downloads calculation): https://crates.io/api/v1/crates/aws-sdk-transcribestreaming/downloads

Braintrust docs sources

• OpenAI integration docs (confirms Transcription as a traced capability class): https://www.braintrust.dev/docs/integrations/ai-providers/openai
• AI providers index: https://www.braintrust.dev/docs/integrations/ai-providers
• AWS Bedrock integration docs: https://www.braintrust.dev/docs/integrations/ai-providers/bedrock

Local repo files inspected

• src/extractors.rs — extract_openai_usage() and extract_anthropic_usage() both operate on token-usage JSON; Transcribe Streaming has no token-usage concept and returns typed event-stream structs, not JSON
• src/stream.rs — BraintrustStream::aggregate()/wrap_stream_with_span() require a Stream<Item = Result<Value, E>> deserializing OpenAI-shaped chunks; incompatible with Transcribe Streaming's bidirectional typed event stream
• src/types.rs — no transcription-oriented span fields (segments, timestamps, confidence, detected language)
• src/lib.rs — public API exports; no Amazon Transcribe references
• Cargo.toml — no aws-sdk-transcribestreaming dependency
• Full local codebase search for transcribestreaming, start_stream_transcription, TranscriptEvent — zero results