FastAligner

FastAligner is a compact English CTC forced-alignment model designed for fast local speech-to-transcript alignment.

It takes known transcript text + audio features and produces CTC logits suitable for forced alignment. It is not intended to perform general-purpose ASR.

Model

• Parameters: 20,615,072
• Architecture: compact LiteConformer-style encoder
• Hidden dimension: 384
• Layers: 6
• Attention heads: 6
• CTC vocabulary: 32
• Input: 80-bin normalized log-Mel spectrogram
• Input rate: 100 Hz
• Output CTC rate: ~50 Hz
• ONNX opset: 18

ONNX variants

Variant	File	Purpose
FP32	onnx/model.onnx	reference model
FP16	onnx/model_fp16.onnx	accelerated inference
INT8	onnx/model_quantized.onnx	compact CPU inference

Inference modes

FastAligner uses one dynamic ONNX model with two runtime modes.

3-second mode

Recommended v0.1 mode.

• 3.0 second chunks
• 0.75 second overlap
• full-utterance feature normalization before chunking
• linear crossfade of overlapping logits
• one global CTC forced alignment after stitching

Full mode

Feeds the complete normalized feature sequence into the same ONNX model without chunking.

See alignment_config.json for the exact runtime parameters.

Preprocessing

Audio must be converted to 16 kHz mono and transformed into an 80-bin log-Mel spectrogram using the settings in preprocessor_config.json.

FastAligner does not accept raw waveform input directly.

Quantization

The repository includes FP32, FP16 and dynamic INT8 ONNX variants.

Timing parity between variants should be validated at the final forced-alignment level rather than judging only raw-logit differences.

Teacher

The student was distilled using facebook/wav2vec2-large-960h-lv60-self as the CTC teacher.

Version

This is FastAligner v0.1.