It is possible it's doing local voice recognition on all audio, and just sending the text in their payloads. If the content is encrypted, there's no way for any network snooping to determine what they're sending.
This is software... saying there's "no way" is usually incorrect.
For example, you could root the device and disable the encryption.
Or, you could perform experiments in a controlled setting where you play a series of identical recordings to Alexa and measure the statistical similarity in the outgoing data for each. The encryption scheme probably provides more entropy than just the audio, so maybe statistical analysis wouldn't help, but it's a start.
