📄 Muslim: A Deployed Arabic Voice AI Platform for Grounded Islamic Knowledge

Summary

Commercial voice assistants handle Arabic weakly and answer religious questions from model memory without citing sources. In this domain that can cause real harm, such as attributing a narration to the wrong source. The paper describes MUSLIM, a deployed voice platform that answers Islamic questions only after retrieving from verified texts.

The browser connects through LiveKit/WebRTC to a stateless agent. The agent runs Silero VAD, NeMo Arabic FastConformer ASR, an OpenAI-compatible LLM and self-hosted TTS, and it calls six MCP servers. Two of these are an exact JSON lookup over eight tafsir books and a self-hosted hadith index. When a user asks for a verse, the system plays a human recording rather than synthesizing it.

The author releases two models, Muslim-6B-PRO and Fasih-TTS-V1. The paper also documents the product layer: per-account turn quotas, a capacity-based refusal, delayed email verification, and monitoring built for a GPU host that fails silently.

Reported results are 98.4% recitation-validation accuracy, 0.9–1.7 s voice latency, and successful retrieval on all 30 held-out questions. The main value is an honest record of what grounded domain QA needs once real users arrive.

Key points

• Retrieval layer: six MCP servers, three run by the operator and three external.
  • Tafsir comes from eight classical books via direct JSON lookup, with no embeddings.
  • Hadith comes from a self-hosted server holding 50,000+ narrations in 17 collections.
• Muslim-6B-PRO: a QLoRA fine-tune (4-bit NF4, r=16) of a 5.94B Qwen3-architecture base.
  • Trained on 2,731 curated examples to route among 31 tools.
  • No benchmark scores are reported yet, and live traffic is still served by a larger general-purpose model.
• Fasih-TTS-V1: a Coqui XTTS v2 fine-tune on about 2.4 h of single-speaker MSA audio.
  • CER of 1.3% on the SILMA benchmark, versus 1.8% for human recordings.
  • Ranked 5th of 17 overall and 2nd of 11 open-weight systems on the Arabic TTS Arena. This is a snapshot from 2026-08-18, based on 245 battles.
• Recitation validator (deterministic, no LLM):
  • Scores 122/124 correct.
  • A census finds that 16.5% of the Qur'an's 6,236 verses share an ambiguous four-word normalized opening with another verse.
• Latency: ASR averages 235.5 ms (RTF 0.025) and LLM first token averages 519.7 ms. Grounding costs only about 50 ms (786 vs 736 ms).
• Access control: the token route refuses with one of six typed reasons. Each account gets 30 turns per rolling 24 h. Tests number over 1,000.

Interesting ideas

• I1. An earlier 99.2% score turned out to come from a run where a missing optional package silently disabled one search layer; a metric that shifts with the installed packages measures the environment, not the system.
• I2. Before blaming the pipeline, the author measured how ambiguous the data itself is (the 16.5% shared verse openings), and the product asks the user to continue instead of guessing.
• I3. When queries name a specific item (a verse, a narration), looking it up directly by its key beats fuzzy embedding-based retrieval on both simplicity and accuracy, and it starts instantly.
• I4. Refusals carry one of a fixed set of machine-readable reasons rather than a generic error, so each failure path can be counted and explained.
• I5. Content that already exists in authentic form (human Qur'an recitation) is never regenerated; it is passed through verbatim.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): every evaluation run should record which components were actually active and fail loudly if one is missing, so scores from witness parsers or repair rules cannot quietly inflate.
• I2 (answer): estimate the share of test items that are genuinely ambiguous, and let readers ask for clarification or abstain on those rather than pick one reading at random.
• I3 (world): resolve compiled knowledge (WordNet synsets, Wikidata ids, VerbNet classes) by exact identifiers wherever possible, keeping every fact traceable to its source entry.
• I4 (answer): give abstentions typed causes, such as no frame found, unresolved coreference, or missing world fact, so error analysis shows which stage blocked the answer.
• I5 (back): when the meaning graph contains quoted speech or fixed expressions, copy the original surface span instead of regenerating it from the graph.