[Ricardo-M-L]

Problem

SpeechSDK.synthesize({ text, audio_setting: { format: 'pcmu_wav' } }) currently sends a 32 kHz sample rate, even though the T2A API documentation describes PCM μ-law output as 8 kHz. The CLI already chooses 8 kHz for both PCM μ-law formats and 24 kHz for Opus when no sample rate is specified, but the SDK always injects 32 kHz.

Change

Use the existing format-aware default in SDK synthesis, for both streaming and non-streaming requests. Preserve explicit sample rates and other caller-provided settings. Make the shared default lookup ignore inherited properties so unknown format strings such as constructor continue to use the numeric fallback rather than causing a missing or nonnumeric sample rate.

The Opus default matches existing CLI behavior; no claim is made that 24 kHz is the only rate accepted by the API. The SDK continues to pass unknown formats through to the API as before.

Validation

• Before the fix, offline mock-request regressions for opus, pcmu_raw, and pcmu_wav sent 32000 rather than the expected defaults; after the initial fix, tests for three inherited-property names reproduced an omitted or nonnumeric rate and pass with the guarded lookup.
• bun test test/sdk/speech.test.ts test/commands/speech/synthesize.test.ts — 65 passed; covers every supported format, unknown/prototype names, explicit options, defaults, and streaming.
• bun run typecheck — passed.
• bun run lint — passed with one existing no-explicit-any warning in the speech test file.
• bun run build and git diff --check — passed.
• Two independent blind reviews of the final diff found no blocking issues. Their focused checks passed (40 and 90 tests respectively).

No live or paid synthesis request was made; tests capture local mock requests and do not verify provider behavior.

AI assistance: Codex investigated, implemented, and tested this change with independent AI-assisted review.

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

[NianJiuZst]

Approve