The PSTN switch-off is not the end of Britain's narrowband problem. It is barely the beginning. Why the UK is rebuilding its telephony backbone around a 1972 codec — and the cost it will impose on every voice AI system in the country for a generation. By Cloudax.
The 2027 PSTN switch-off is sold as Britain stepping into a modern voice future. In practice, the country is rebuilding its telephony backbone around a 1972-era codec — G.711 — and shipping every voice AI system through a 3.4 kHz audio channel that the rest of the world quietly moved past a decade ago.
The mobile carriers, hyperscalers and large enterprise communications platforms outside the UK have been steadily moving to wideband and super-wideband codecs — Opus, AMR-WB, EVS — for years. These deliver 16 kHz or 32 kHz audio, the bandwidth modern speech recognition and speech synthesis stacks were trained to operate inside. The UK's all-IP replacement for the PSTN has largely kept G.711 (8 kHz, 64 kbit/s) as the dominant signalling and media codec for interoperability reasons.
State-of-the-art ASR and TTS systems are trained on wideband audio. When their inputs are downsampled to 8 kHz telephony, recognition accuracy drops, voice quality flattens, and the perceptual gap between an AI voice and a human voice widens substantially. The same models that pass for human on a wideband call sound noticeably synthetic on a UK landline. For voice AI vendors trying to clear the uncanny valley, the British network adds an extra meter to the jump.
Beyond raw audio quality, narrowband telephony interacts badly with modern voice AI in ways that show up as production failures: degraded turn detection in noisy conditions, accent-specific recognition loss, mispronunciation of numerical sequences, brittle DTMF handling on certain SIP trunks, packet loss concealment artifacts that read as confidence to the model, and audio cropping at the start of utterances that loses the first phoneme of a reply. None of these are addressed by the switch-off plan.
Three things would meaningfully shift the trajectory: a wideband codec floor for all-IP voice calls between UK operators, mandatory minimum jitter and packet loss budgets on the access network, and a published quality-of-service contract for AI-mediated calls. Without those, every voice AI system built in Britain will operate at a structural quality disadvantage to its international counterparts for a generation.