📄 MyVoice: Arabic Speech Resource Collaboration Platform

Summary

Dialectal Arabic is poorly served by speech technology, mainly because large, openly licensed recordings covering its many regional varieties are scarce and expensive to produce. MyVoice is a web-based crowdsourcing platform built to address this gap. Volunteers pick a fine-grained dialect, down to the city or country level, and read aloud prompts shown on screen. The same people can also act as validators, rating and annotating recordings made by others. Before any human sees a recording, an automatic quality layer screens it. This layer combines voice activity detection with a multi-dialect Arabic ASR system that assigns each submission a confidence score. Administrators can upload new prompt sets, define tasks that go beyond dialect speech collection, monitor statistics, and remove poor recordings or block abusive users. The paper is a short system description and reports no dataset sizes or experiments. Its contribution is reusable infrastructure for community-driven, open collection of under-represented speech data.

Key points

• The platform is available at myvoice.arabicspeech.org and was developed at the Qatar Computing Research Institute (HBKU).
• The front-end uses Nuxt3 and TailwindCSS. The back-end uses FastAPI served by Uvicorn, Supabase for authentication and user metadata, and PM2 for process management, restarts and logs.
• Quality assurance relies on voice activity detection plus a confidence score from a multi-dialect, code-switching-aware Arabic ASR model (Chowdhury et al., 2021). Admins use this score to find outliers and low-quality submissions.
• The interface has five views:
  • Tasks: contributors choose a dialect and record prompts.
  • Validation: contributors review their own recordings and redo weak ones.
  • Submissions: admins inspect each recording and its score.
  • Users: admins inspect individual contributors and can block them.
  • Admin: admins view statistics over a chosen time range, upload prompts and grant admin rights.
• No collected-data statistics or evaluation results are reported.

Interesting ideas

• I1. Running an automatic model-confidence filter before human validation means reviewers only spend effort on recordings that pass a cheap machine check. — For Mova (learn): Score silver trees by how strongly the witness parsers agree, and send only borderline cases to the LLM teacher tiers or a human, which saves expensive review effort.
• I2. Each contribution is tagged with fine-grained variety metadata (city or country dialect) at collection time rather than inferred later. — For Mova (tree): Attach source and dialect or genre labels to every treebank sentence, so that parser accuracy and UD dialect conversion can be measured per stratum instead of as one blended average.
• I3. Contributors can switch between producing data and validating others' data, and their feedback is then checked by administrators, giving a layered trust chain. — For Mova (learn): Mirror this chain so that one witness parser proposes, a different witness or teacher tier validates, and a final arbiter audits disagreements, keeping every acceptance step explainable.
• I4. Every submission keeps its contributor's identity, so a contributor found to be unreliable can be blocked and all of their past data removed in one action. — For Mova (learn): Record which parser, teacher tier or repair rule produced each silver annotation, so a component later shown to be faulty can have all its outputs retracted at once.
• I5. The platform is built around a generic task abstraction, which lets admins add new kinds of collection jobs without changing the system. — For Mova (answer): A similar pluggable task queue could host new test families (school tests, word problems, fable QA) and route items to the right reader without code changes.