Hume introduces an API for questions about speech and sound Hume introduces an API for questions about speech and sound

Questions replace fixed labels

Developers supply the question and its possible answers. Supported forms cover yes-or-no judgments, multiple choices, ordered scales and the timing of sounds. Hume describes a workflow in which a voice agent combines these judgments with a separate transcript and its business rules.

Probabilities describe listeners

The Python SDK defines a probability as the estimated share of human raters who would choose an answer. A value around 0.5 represents divided listeners, not an uncertain model. All questions in a request share one model inference. Adding, removing or reordering them can change an answer. Each response associates answers with the question identifiers and includes a run identifier. Run-history endpoints expose the status and original request, but do not recover the answers. An application that needs an audit trail must therefore retain the returned results itself.

Long calls become separate inputs

The Python SDK can resample and mix supported WAV recordings before upload, matching the service’s single-channel, 16-kilohertz input. It also offers a first-channel option for recordings that isolate a speaker there. Longer recordings are divided into separately billed requests. Each part receives the same questions and context, but the model hears only that part; the SDK does not combine their judgments. Its default even split avoids a short final remainder. A 32-minute recording becomes seven roughly equal pieces at the maximum five-minute chunk setting. MP3 cuts can repeat frames needed for decoding, so an event near a boundary can appear in two parts.

Partial completion is visible

The SDK records success or failure separately for each chunk. Once one part has succeeded, an audio-decoding failure can be isolated to a later damaged section while processing continues. Other failures stop the job and leave subsequent parts marked as not attempted. The completion flag is true only when every requested part succeeds. A resume position identifies where to restart after a stopping error, but earlier decoding failures are not retried by that operation. This lets applications distinguish a complete review from a successful response covering only part of a call.

A different role from voice generation

The launch follows Hume’s October 2 notice that its text-to-speech and Empathic Voice Interface APIs will end access on November 13 at 12:01 a.m. Eastern Standard Time. Those services remain supported until that deadline. Audio judgments occupy a different part of an application from generating a spoken answer. Adopting this new service does not replace a voice-generation integration affected by that earlier notice.