Persian text-to-speech research for MCI

Persian text-to-speech research, commissioned and supervised by MCI (Hamrah-e Aval) under Iran's National Elites Foundation. The hard part is linguistic: Persian script leaves short vowels and the ezāfe unwritten, so pronunciation has to be decided from context.

At a glance

  • Persian script leaves short vowels and the ezāfe unwritten, so pronunciation is ambiguous without context.
  • Persian text-to-speech research: a pipeline from text through pronunciation, synthesis and a vocoder to speech.
  • Completed, June 2021 to September 2022, under MCI's supervision and the National Elites Foundation.

Highlights

  • Pronunciation stage

    Grapheme-to-phoneme conversion decides how written Persian is pronounced.

  • What Persian script leaves out

    Short vowels and the ezāfe are not written, and homographs share one spelling, so pronunciation is ambiguous without context.

  • Synthesis model and vocoder

    A text-to-speech model and a vocoder then produce Persian speech.

The challenge

Persian text-to-speech is materially harder than English, because converting written letters into sounds is ambiguous without context. The work had to contend with three properties of Persian script:

  • Short vowels are not written, so the pronunciation is underdetermined by the text alone.

  • The ezāfe construction is not written and has to be inferred.

  • Homographs share one spelling across different words and cannot be told apart without context.

Architecture

A Persian text-to-speech pipeline: written text passes through grapheme-to-phoneme conversion, which decides how it is pronounced, and a text-to-speech model and a vocoder then produce Persian speech.

What I built

Text-to-speech for Persian, delivered under national-operator supervision. The work was commissioned and supervised by MCI (Hamrah-e Aval), Iran's largest mobile operator, and managed by the National Elites Foundation under its Shahid Ahmadi Roshan programme. It is a national-scale project, with a named operator and a national foundation behind it.

Outcome

The research was delivered under the supervision of MCI (Hamrah-e Aval), within the National Elites Foundation's Shahid Ahmadi Roshan programme. The engagement ran from June 2021 to September 2022.

My role

  • AI Researcher on Persian text-to-speech within the National Elites Foundation's Shahid Ahmadi Roshan programme.
  • Carried out the work under the commission and supervision of MCI (Hamrah-e Aval).

Outcomes

  • Persian text-to-speech research delivered under the supervision of MCI (Hamrah-e Aval).
  • Managed by Iran's National Elites Foundation within its Shahid Ahmadi Roshan programme.
  • Engagement completed: June 2021 to September 2022.