Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Malayalam Language Support

Status: Current Last updated: 2026-08-06 16:10 EDT

Malayalam (mal) is supported in transcribe-only mode. Stanza ships no processor packages for Malayalam, so utseg / morphotag / forced alignment via Stanza are unavailable. ASR runs via the whisper_hub engine using a HuggingFace Whisper fine-tune.

This page covers Malayalam concretely; the same pattern applies to the other Indo-Aryan / Dravidian / Austronesian languages whose entries in Stanza’s resources.json are charlm/lang_name stubs without processor packages, see the Stanza capability table at worker startup for the authoritative current list.

Quick Reference

Pipeline StageMalayalam-Specific Behavior
ASRHuggingFace Whisper fine-tune via --asr-engine whisper_hub
Text normalizationNone, Malayalam script passed through as-is
Number expansionRust-side NUM2LANG table. Digits like "3" reach CHAT as "മൂന്ന്" and pass E220. See Number Expansion for the dispatch path; Malayalam is registered in the per-language coverage matrix on that page.
RetokenizeNot applied, retokenize maps a Stanza-tokenized word list back to ASR tokens; without a Stanza pipeline there is nothing to map.
MorphosyntaxNot available: Stanza ships no Malayalam pipeline
UtsegNot available: same reason
Forced alignmentStock Wave2Vec MMS works on the audio; no Malayalam-specific preprocessing

Running Transcribe

batchalign3 transcribe --lang mal \
    --asr-engine whisper_hub \
    audio_dir/ -o output_dir/

The whisper_hub engine routes to whichever HuggingFace Whisper fine-tune the model resolver picks for mal. The output CHAT file contains a single main tier with the transcribed Malayalam text and a media reference; no %mor, no %utseg, no per-word timing.

thennal/whisper-medium-ml produces clean output on the test sample the Whisper-Hub integration was validated against. Other Malayalam fine-tunes exist on HuggingFace; the resolver in batchalign/models/resolve.py accepts an explicit override if a different checkpoint is preferred.

Why no morphotag / utseg

Stanza’s resources.json contains a Malayalam entry, but only with backward_charlm, forward_charlm, and lang_name fields, no packages key listing tokenizer / POS / lemma / depparse models. The Python worker’s UnsupportedLanguageError preflight catches this before stanza.Pipeline() runs and the optional sub-stages in transcribe are dropped from the plan at job submission.

Adding morphotag / utseg for Malayalam would require either Stanza upstream shipping Malayalam pipelines, or training and integrating a non-Stanza tagger, neither is in scope.

Why a fine-tune instead of stock Whisper

Stock multilingual Whisper does include a Malayalam language ID and can transcribe Malayalam audio, but accuracy on extended Malayalam recordings is generally well below what a Malayalam-specific fine-tune delivers. The whisper_hub engine exists specifically to route audio to such fine-tunes when the user requests it.

For languages where stock Whisper is already strong (English, Spanish, etc.), whisper_hub is unnecessary, the default --asr-engine whisper is fine.

This transcribe-only mode applies to other languages where Stanza has no processor packages. Check at worker startup which codes the capability table reports as supported; codes outside that set follow the same Malayalam pattern (transcribe via stock Whisper or a fine-tune; no Stanza-driven analysis).

For context on Indo-Aryan / Dravidian languages that do have full Stanza support, Tamil (tam), Hindi (hin), Urdu (urd), Telugu (tel), and Thai (tha) ship complete processor packages in current Stanza, morphotag and utseg work the standard way, no whisper_hub needed. The exact set varies by Stanza version; the worker’s capability table at startup is authoritative.

Resolved issues

E220 on Whisper digit emissions (resolved)

HuggingFace Whisper fine-tunes for Malayalam (including thennal/whisper-medium-ml) transcribe spoken numbers as Arabic digits (“3”, “100”) rather than Malayalam script. Pre-fix, CHAT validation rejected these with E220 because mal is not in the digit-allowed language allowlist (talkbank-tools/../chatter/crates/talkbank-model/src/validation/word/language/digits.rs, which permits digits only for zho, cym, vie, tha, nan, yue, min, hak), and the Python num2words library has no Malayalam backend.

Fix: added a Malayalam entry to NUM2LANG in crates/batchalign-transform/data/num2lang.json covering 0-20, decades 30-90, plus 100/1000 anchor words. The per-word Rust pass in crates/batchalign/src/pipeline/transcribe.rs:527::prepare_asr_chunks calls expand_number(text, "mal") on every word, converting digits to their Malayalam-script word forms. The Python IPC path no longer exists, Malayalam expansion is end-to-end Rust. Tests at crates/batchalign-transform/src/asr_postprocess/num2text.rs:540 (malayalam_single_digits_expand_to_script, malayalam_digits_collected_for_expansion, malayalam_anchor_decades_and_hundreds) lock in the expected expansions.

Higher-magnitude numbers (4-digit and beyond) are decomposed by decompose_with_table greedily against the anchor entries; if the table can’t fully decompose, the original digit string is returned (matching every other language without an exhaustive table).

Operational notes

  • Audio files of arbitrary length are supported; ASR processes in 25-second chunks via the HuggingFace pipeline (batchalign/inference/asr.py:177 chunk_length_s=25).
  • Output CHAT can be edited by hand and re-run through other TalkBank tooling; the absence of %mor / %utseg does not affect downstream tools that don’t require them.

This page last changed: 2026-08-06 (commit ff0d5869). The whole book last changed: 2026-09-16 (commit 34d249d8).