MFA Quran Hafs
Forced alignment for recitation in the riwaya of Hafs. It times every phone of an ayah against its text, so madd, ghunnah and qalqalah can be measured, not only word boundaries.
- Task
- Forced alignment
- Riwaya
- Hafs 'an 'Asim
- Phones
- 67, one per articulation
- Dictionary
- 20,967 entries, phonetized from whole ayat
- Rule index
- 93,430 tajweed positions
- Training audio
- About 110 hours, balanced by style, 40 of them mujawwad
- Built with
- Montreal Forced Aligner 3.4
- Licence
- Apache-2.0, open download
See it work
Al-Fatiha with its words and sounds drawn to scale on one time axis. A madd is as wide as it was held.
Why recitation needs its own aligner
General Arabic aligners fail on recitation. Madd vowels held for seconds, ghunnah, the melody of mujawwad and the reverberation of a mosque are far outside ordinary speech. This model was trained and measured on exactly those.
Two choices matter most. The dictionary phonetizes each ayah as a whole, with the joined forms and the idgham, ikhfa and iqlab that cross word boundaries; forms taken from isolated words measure 20.5% phone error, contextual ones 1.6%. And every articulation is one phone: a long vowel of any prescribed length is a single symbol, not a repeated one, which an aligner of this kind handles badly.
The training audio is balanced by style: ayah clips from QUL reciters checked against a human word count, extra examples of lazim and muttasil madd, and 40 hours of mujawwad. A corpus five times larger but unbalanced made mujawwad coverage worse, and a mujawwad-only model did worse than this one.
Measured
From the model card.
| Measure | Result |
|---|---|
| Phone boundaries against landmarks in the signal (533 geminate stop releases, 4 reciters including mujawwad) | 10 to 20 ms median |
| Speech covered by no phone, murattal, reciters held out | 0.0 to 0.4% |
| Speech covered by no phone, mujawwad, a reciter held out | 5.0% |
| Madd length against the prescription, tempo-normalized: natural (2) and munfasil (4) | 2.1 / 4.8 |
| Word boundary on a human-labelled mujawwad elongation | within about 100 ms |
Use it
One 16 kHz mono wav and one .lab file per ayah clip. The .lab holds the Uthmani text of the ayah in whole words; the dictionary does the phonetization.
mfa align CORPUS_DIR quran_hafs.dict quran_hafs_acoustic.zip OUT_DIR \
--beam 40 --retry_beam 160Keep every path in plain ASCII: OpenFST on Windows fails on other characters.
| File | What it is |
|---|---|
quran_hafs_acoustic.zip | The MFA acoustic model |
quran_hafs.dict | Pronunciation dictionary, 20,967 contextual entries |
tokens.txt | Phone symbol table, 67 phones and a blank, for CTC integrations |
rule_index.jsonl | Tajweed rule annotations per ayah: 93,430 positions, each with its governing rule and prescribed length |
Measuring tajweed with it
- 1
QUL's word timings start about 245 ms early against the physical start of the sound, a playback convention. Compare against landmarks in the signal instead.
- 2
Trust spans that begin and end on obstruents. Boundaries inside a run of sonorants are less certain, so measure a rule's segment between obstruent anchors.
- 3
Before a pause, breath and the decay of the room attach to the last phone. Trim the end of a madd by energy and voicing before measuring it.
Known limits
Measured, not hypothetical. From the model card.
- 1
Hafs only.
The phonetizer supports no other riwaya. Aligning Warsh or Qalun audio with it would mislabel exactly the features tajweed cares about.
- 2
A stop in the middle of an ayah is not modelled.
The Uthmani text carries no waqf marks, and a reciter who stops mid-ayah uses pausal forms the dictionary does not offer.
- 3
Some sung sound counts as silence.
In mujawwad, about 5% of the voice, mostly the fading end before a breath, is attributed to silence. Use trims on the signal side for duration work there.