S2S · Forensics

What is actually inside the audio?

Codec history fingerprints the delivered audio for the block-boundary periodicity a lossy codec (Opus, MP3, AAC) leaves behind, and compares what the stream declares against what it actually contains. For a native S2S model the question is whether its output silently carries a fingerprint of the caller's input codec — an echo of your pipeline rather than its own clean stream.

Preliminary: a single native S2S model, measured across n=3 caller voices. A mismatch flags a stream worth inspecting, not proof of vendor-side re-encoding — a fingerprint can originate anywhere in the path.

Codec-History Contamination

match = no echo · bettern=3 voices (median)

Plain English: does the model's audio carry a fingerprint of the caller's audio codec, instead of its own? A mismatch means the model is echoing your input pipeline.

ProviderDeclaredDetectedConfidenceMatchNotes
native S2S modelpcmpcm1.00matchPCM end-to-end via Speko WS

What we found: the model declares and emits PCM end to end — declared and detected codecs match, so no input-codec echo. Rows where declared ≠ detected would mean the model is silently re-encoding or passing input frames through.