For developers

Audio preparation

Every recording goes through the same steps before it is scored, with the same limits for every caller. They are Spelly’s standard and cannot be changed per call or per key.

#What you can send

Any common audio format is decoded: WAV, FLAC, OGG, WebM, MP3 and others. WAV at 16 kHz, 16-bit, mono needs no conversion. Stereo is mixed down to mono, and every recording is resampled to 16 kHz.

#Length and size

An upload larger than 960,000 bytes is refused with HTTP 413 (assess.recordingTooLong). Only the first 15 seconds of a recording are scored.

#Silence at the edges

Silence before the first and after the last speech is cut, keeping 750 ms around the speech. Speech is anything louder than 5% of the recording’s loudest level held for 150 ms, so a click at the edge cannot set the level.

#Recordings that are refused

A refused recording gets no score and costs nothing: its charge is given back.

  • Unreadable (HTTP 400, assessment.audioUnreadable): no decoder can read the file.
  • Nothing heard (HTTP 422, assessment.nothingHeard): the recording is silent (its peak is below 0.01 of full scale, or its loudest level is below -50 dBFS), it is noise rather than speech (a mean zero-crossing rate above 0.25 while its loudest level stays below -35 dBFS), or the model hears no sound in it.
  • Too short (HTTP 422, assessment.tooShort): shorter than 85 ms, the least the model reads, or too short to hold its text: the model needs one frame of about 40 ms for each expected sound, and one more between two identical sounds in a row.

Questions?

Tell us what you have in mind and we will get back to you.

Get in touch