sk-align

Standalone forced alignment for Scottish Gaelic — no Kaldi or PyKaldi dependency.

sk-align reimplements Kaldi's nnet3 forced-alignment pipeline entirely in Python/NumPy/PyTorch, reading Kaldi model files directly. It produces word-level timestamps at parity with PyKaldi while being easier to install and deploy.

---

Features

• Zero Kaldi dependency — pure Python reads Kaldi binary formats (final.mdl, tree, L.fst, etc.)
• from_pretrained() — one-line model download from Hugging Face Hub
• MFCC extraction — vectorised NumPy implementation matching Kaldi output
• TDNN-F nnet3 inference — full PyTorch reimplementation of the forward pass
• k2 Viterbi decoder — fast FSA-based decoding via intersect_dense + shortest_path
• Word-level timestamps — [{"word": "hello", "start": 0.12, "end": 0.45}, ...]
• Parity-tested — 55 tests verify numerical match against PyKaldi reference

Installation

pip install sk-align              # core (numpy + scipy + torch + k2)
pip install sk-align[all]         # + huggingface_hub for from_pretrained()

Or install from source:

git clone https://github.com/your-org/sk-align.git
cd sk-align/sk-align
pip install -e ".[all]"           # editable with all extras

Optional extras

Extra	Installs	Needed for
hub	huggingface_hub>=0.20	Aligner.from_pretrained()
all	huggingface_hub	Full end-to-end pipeline
test	pytest + huggingface_hub	Running the test suite
dev	test extras + ruff	Development

Quick start

from sk_align import Aligner

# Download model from Hugging Face and load (cached after first call)
aligner = Aligner.from_pretrained()

# audio: float32 numpy array, 16 kHz, mono
timestamps = aligner.align(audio, ["cumaidh", "sinn", "a'", "dol"])
# [{"word": "cumaidh", "start": 0.33, "end": 0.72},
#  {"word": "sinn",    "start": 0.72, "end": 0.99},
#  ...]

Loading a local model

from sk_align import Aligner
from sk_align.nnet3_torch import TorchNnetScorer

scorer = TorchNnetScorer.from_model_file("/path/to/model/final.mdl")
aligner = Aligner.from_model_dir("/path/to/model", nnet_scorer=scorer)

timestamps = aligner.align(audio, words)

Using pre-computed log-likelihoods

import numpy as np
from sk_align import Aligner

aligner = Aligner.from_model_dir("/path/to/model")  # no scorer needed
loglikes = np.load("loglikes.npy")  # (num_frames, num_pdfs)

timestamps = aligner.align_with_loglikes(loglikes, words)

Architecture

The alignment pipeline reimplements each stage of Kaldi's forced alignment in pure Python:

Audio (float32, 16 kHz)
  │
  ▼
┌─────────────────────┐
│  MFCC Extraction    │  sk_align.mfcc        (NumPy, batch-vectorised)
└─────────┬───────────┘
          ▼
┌─────────────────────┐
│  Nnet3 Forward Pass │  sk_align.nnet3_torch  (PyTorch TDNN-F)
└─────────┬───────────┘
          ▼
┌─────────────────────┐
│  Graph Compilation  │  sk_align.graph        (L ∘ G, context expansion)
└─────────┬───────────┘
          ▼
┌─────────────────────┐
│  Viterbi Decoding   │  sk_align.k2_decoder   (k2 FSA intersection)
└─────────┬───────────┘
          ▼
┌─────────────────────┐
│  Word Alignment     │  sk_align.word_align   (boundary extraction)
└─────────────────────┘
          │
          ▼
  [{"word": "...", "start": 0.12, "end": 0.45}, ...]

Modules

Module	Description
sk_align.aligner	High-level Aligner class — main entry point
sk_align.mfcc	MFCC feature extraction (batch NumPy, Kaldi-compatible)
sk_align.nnet3_model	Kaldi nnet3 binary parser
sk_align.nnet3_torch	PyTorch reimplementation of TDNN-F forward pass
sk_align.fst	OpenFst binary format reader + FST representation
sk_align.graph	Per-utterance decoding graph compiler (L ∘ G + context expansion)
sk_align.tree	Kaldi ContextDependency tree reader
sk_align.transition_model	Kaldi TransitionModel reader
sk_align.k2_decoder	k2-based Viterbi decoder
sk_align.word_align	Word boundary extraction + timestamp conversion
sk_align.kaldi_io	Low-level Kaldi binary I/O helpers

Model

The default model is hosted at eist-edinburgh/nnet3_alignment_model on Hugging Face Hub. It is a TDNN-F nnet3 alignment model (3456 PDFs) trained for Scottish Gaelic.

Expected model files:

final.mdl           TransitionModel + nnet3 weights
tree                ContextDependency tree
L.fst               Lexicon FST (OpenFst binary)
words.txt           Word symbol table
disambig.int        Disambiguation symbol IDs
word_boundary.int   Phone word-boundary types

Testing

The test suite verifies numerical parity with PyKaldi at every stage.

pip install -e ".[test]"
pytest                   # 49 tests — MFCC, I/O, graph, decoder, end-to-end parity

Tests include:

• MFCC parity — feature output matches Kaldi within floating-point tolerance
• I/O round-trip — all Kaldi binary readers produce correct data structures
• Graph compilation — decoding graphs match expected state/arc counts
• Decoder parity — k2 decoder alignment matches reference Viterbi output
• End-to-end parity — word timestamps match PyKaldi within 30ms

Performance

Benchmark on a 5-second Scottish Gaelic utterance (25 words), CPU:

Stage	Time	% of total
MFCC	25 ms	4%
Nnet3 forward	434 ms	75%
Graph compile	46 ms	8%
k2 decode	72 ms	13%
Word align	<1 ms	<1%
Total	578 ms	—

End-to-end throughput is at parity with PyKaldi (~560 ms per utterance).

License

MIT