I ported Meta's MMS forced aligner to TypeScript: word-level timestamps in Node.js, 42% faster than Python

Description and use

The mms-forced-align npm package implements Meta's MMS forced alignment in Node.js, achieving 42 % faster inference and 32 % less memory than Python while matching reference timings exactly.

Problem

Need word-level timestamps for captions in a Node.js service without running a heavy Python sidecar.

Outcome

The TypeScript port matches Python outputs exactly, runs 42 % faster, uses 32 % less memory, and supports 17 Indic languages via a romanizer.

Workflow

• 1. Accept mono 16 kHz Float32Array audio
• 2. Normalize audio (zero mean, unit variance)
• 3. Split into 20 ms frames (320 samples)
• 4. Run ONNX model to get logits per frame
• 5. Apply log_softmax to obtain log-probabilities
• 6. Convert transcript words to extended CTC sequence with blanks
• 7. Run custom TypeScript CTC Viterbi decoder over the grid
• 8. Map Viterbi path to letter and word spans in seconds

Evidence

mms-forced-align runs Meta's MMS forced-aligner model in Node.js through ONNX Runtime, with a CTC Viterbi decoder I wrote from scratch in TypeScript (~170 lines).

Source

~42% faster than Python's torchaudio on the same fp32 model, with ~32% less peak memory.

Source

37 of 37 reference words match torchaudio to 0.0 ms. On a real 116-word clip, 110 words are identical.

Source