The most expensive speech recognition model cannot correct an audio stream decoded at the wrong sample rate. Encoding, audio framing and timestamps deserve their own acceptance tests.
Sample rate identifies samples, not quality by itself
An 8 kHz telephone stream carries a narrower audio frequency band than a clean 16 kHz capture. Resampling a telephone stream upward does not recreate missing high frequencies. Conversely, sending 16 kHz samples while claiming 8 kHz can double apparent duration or distort words. Keep capture sample rate, resampler output and API declaration explicit. A 16-bit sample value is not the same thing as 16 kHz, and mono is not automatically equivalent to averaging stereo without checking phase and speech content.
The wire format can cause silent failures
PCM is uncompressed sample data, while MP3, Opus and Speex are coded formats with framing and decoding rules. A WebSocket text frame containing base64 audio is not a raw binary PCM buffer. Providers publish exact accepted combinations for each endpoint; one service’s Opus setting should not be copied into another. Before launch, feed short test clips with known duration and inspect the decoded length, waveform range and expected transcript. Reject unsupported encodings explicitly.
Understand realtime packet cadence
Tencent’s published realtime ASR guide suggests sending about 200 milliseconds of audio per 200 milliseconds of real time. That prevents an application from dumping the whole file into a live socket while expecting stable incremental hypotheses. Network jitter, device buffering and client reconnects can all affect delay. Keep packet timestamps and a bounded ring buffer, detect gaps, and avoid replaying already finalized segments after recovery. End-to-end latency includes playback or display, not only server compute.
Integrate speech generation and listening carefully
An agent may receive microphone PCM, convert it for recognition, then output generated voice in a different format. Always track codecs at each stage, since a speech generator may return OGG/Opus, MP3 or PCM depending on its interface. Decode to a consistent mixer format before combining with background audio, and test echo cancellation and barge-in. Resampling and format conversion should be recorded in the test protocol, because they materially affect perceived quality and turnaround.
Build the acceptance matrix
Test sample-rate mismatches, invalid headers, silence, clipped audio, channel swapping, high packet jitter, long pauses and unusually loud input. Report how often users hear first output, how many packets are dropped, how many transcript words are corrected later and whether each endpoint returns actionable errors. Documented audio-format support is not an independent claim about speech quality; actual user studies still have to be performed.
Evidence, availability and commercial-rights boundary
Research cutoff: 25 October 2026. Provider documentation is evidence of its published interface or policy, not independent product testing. No performance score, hands-on trial, customer consent, territory-wide entitlement or legal clearance is claimed. Confirm the exact version, selected region, voice rights, data flow, account limits and current service terms before deployment. Unknown model release dates, account prices and plan limits are deliberately left unasserted; monetary claims belong to the separately dated AI10 pricing ledger. Where documentation is undated or dynamic, the public page is not treated as an archival snapshot of every earlier tier.