Switch-Point Prosody: Why AI Voices Break Where Pakistani Ads Switch Languages

A research proposal by Muhammad Usman, BS Artificial Intelligence, Superior University, Lahore — October 2026

Walk through any commercial break on Pakistani television, or scroll through sponsored video ads on Instagram and TikTok, and a linguistic pattern repeats: the pitch is neither pure Urdu nor pure English. It is a deliberate mix — Urdu as the matrix language carrying the message, English insertions carrying prestige, urgency, and the call to action:

"Yeh phone best hai — limited offer, abhi order karo!" "Naya smart TV sirf Rs. 85,000 me — free delivery ke saath!"

This is not sloppy speech. In Myers-Scotton's Matrix Language Frame model, bilingual speakers systematically embed one language's content morphemes inside another language's grammatical frame, and advertisers exploit this: English words signal modernity and trigger the purchase impulse, while Urdu carries trust and cultural proximity. The persuasive force of the ad lives exactly at the switch points — the emphatic stress on "best", the rising pitch into "order now!"

AI-generated voiceovers are now cheap enough for Pakistan's small businesses to use in video ads. But every practitioner who has tried knows the failure: feed a code-switched script to a TTS engine and the voice stumbles precisely where the ad needs it to soar. The Urdu words come out with English phonemes, the stress lands on the wrong syllable, and a flat pitch contour replaces the salesman's rise.

I verified this with a live demonstration: synthesizing "Yeh product best hai, order now" through a production English TTS voice, the Urdu function words receive English phonemization, stress defaults to English patterns, and the persuasive pitch rise into the call to action is absent.

Code-mixed speech synthesis is an established field — but for other language pairs. Sitaram and Black (CMU) built an early Hindi-English framework and found significant listener preference for systems that correctly identify and pronounce words in each language. Joshi and Garera (2023) reached production quality for Hindi-English e-commerce voice bots with a positive comparative MOS against a commercial baseline. Mandarin-English bilingual synthesis was demonstrated at Interspeech 2020, and an ICASSP 2020 system explicitly predicted log-F0 because phonetic representations alone lack the prosody code-switched synthesis needs.

Crucially, the prosodic disruption at switch points has been measured: a Chinese-English bilingual speech study found English words inside code-switched utterances showed nearly double the pitch variability of monolingual speech — average F0 standard deviation of 51.9 Hz versus 29.8 Hz. Speakers measurably alter their prosody around switch points, and a TTS system trained on monolingual prosody misrenders exactly those regions.

The gap: no published work addresses Urdu-English switch-point prosody, and none targets advertising persuasion — despite Urdu-English mixing being the everyday reality of hundreds of millions of speakers.

Phase 1 — Corpus: 50 code-switched Pakistani ad scripts, 5 native bilingual speakers, switch points annotated. Phase 2 — Acoustic measurement: F0 contours via pYAAPT; compare 300ms windows around switch points against monolingual windows; synthesize the same scripts with commercial TTS engines and repeat. Phase 3 — Perceptual evaluation: ABX listening tests (n ≥ 40): naturalness (MOS) and purchase intent ratings.

Expected contributions: the first acoustic characterization of Urdu-English switch-point prosody, a public annotated corpus, quantified human–synthetic prosody gaps, and a prosody-aware synthesis recipe evaluated on purchase intent — the metric advertisers actually care about.

This is a research proposal — the study has not yet been executed. Findings from Hindi-English and Mandarin-English systems may partially transfer, which would narrow but not eliminate the novelty. All literature claims are verifiable through the references in the full paper.