WebRTC and why audio quality decides AI accuracy

Every discussion of speech recognition accuracy assumes the model is the variable. Frequently it is not. The audio arriving at the model has already been band-limited, compressed and, on a poor network, partially discarded. No model recovers information that was never transmitted.

What narrowband telephony removes

Traditional telephony carries roughly 300Hz to 3.4kHz, sampled at 8kHz. Human speech has meaningful content well above that. The consonants that distinguish similar words — s, f, th, sh — live largely in the higher frequencies that narrowband discards.

The practical consequence is that word error rate on narrowband audio is materially worse than the same model on wideband audio of the same speech, and the errors concentrate on exactly the words that change meaning — names, reference numbers, yes versus no in noise.

Why WebRTC rather than a codec upgrade alone

Opus at 48kHz is part of it, but the transport matters as much as the codec. WebRTC was designed for real-time media on unreliable networks, and brings several things that matter directly to a voice agent:

• Opus adapts bitrate to available bandwidth instead of degrading abruptly
• Forward error correction reconstructs lost packets without waiting for retransmission
• Jitter buffers adapt continuously rather than being fixed
• Echo cancellation and noise suppression run before the audio reaches transcription
• Direct peer connection avoids the transcoding hops that each cost quality and latency

Packet loss is an accuracy problem, not a comfort problem

A human listener interpolates through brief dropouts without noticing. A transcription model does not: a lost packet is missing audio, and the model produces its best guess at what was said in a window that has a hole in it. This is why calls that sound acceptable to a person can transcribe badly.

The same applies in the other direction

Synthesised speech is generated at high sample rates and then has to survive the transport to the customer. Downsampling a natural-sounding voice to narrowband undoes much of what makes it sound human — the same high-frequency detail that carries consonant clarity also carries the texture people hear as natural. A wideband path preserves the quality you paid for in the TTS engine.