Experiments — Lab Scaffolding

This directory contains the lab scaffolding for the Speech Recognition Course: working programs with the key pieces left out for you to implement. Each unfinished script has a TODO(M#) block and raises NotImplementedError until you complete it.

[!TIP] Stuck, or want to check your work? The complete versions of every script live in ../Solutions/. Finish the lab first, then compare.

How the labs work

  1. Open the scaffolding script for the module you’re on (see the table below).
  2. Read the header comment and the TODO(M#) block — they describe exactly what to implement and which helper functions are already provided.
  3. Fill in the missing code.
  4. Run it (commands below). Compare your output against the reference in ../Solutions/.

Each module’s README.md (../M1_Introduction/, ../M3_Acoustic_Modeling/, etc.) has the “Lab for Module N” write-up with the full background.

Which scripts have exercises

Script Module Your task
M1_Score.py M1 Implement score() — compute WER and SER from a ref/hyp TRN pair
M2_Wav2Feat_Single.py M2 Complete single-file feature extraction
M3_Train_AM.py M3 Implement the training and evaluation loop in train_model()
M3_Plot_Training.py M3 Parse a training log and plot the curves

The remaining files are provided complete as supporting code and are identical to the versions in ../Solutions/: M2_Wav2Feat_Batch.py, arpa2fsa.py, StaticDecoder.py, and the helper modules speech_sigproc.py, htk_featio.py, wer.py.

[!NOTE] am/, lists/, and misc/ here are symlinks into ../Solutions/, so the data is shared and code that reads am/, lists/, or misc/ relative to this directory works unchanged. A git clone on Linux/macOS reproduces the symlinks automatically; on Windows you may need Developer Mode / symlink support enabled.

Prerequisites

Dependencies are managed with uv. From the repository root:

uv sync

This installs numpy, scipy, matplotlib, soundfile, and a CPU-only torch into .venv. Run scripts with ../.venv/bin/python (from this directory).

[!IMPORTANT] M2 feature extraction and everything downstream need the LibriSpeech corpus, which is not bundled. The M1 lab uses the sample transcripts in misc/ and needs no corpus.

Running your work

Run all commands from this Experiments/ directory.

# Module 1 — score the sample transcripts (no corpus needed)
../.venv/bin/python M1_Score.py -rt misc/ref.trn -ht misc/hyp.trn

# Module 2 — feature extraction (needs LibriSpeech)
../.venv/bin/python M2_Wav2Feat_Single.py

# Module 3 — train an acoustic model, capturing the log for plotting
../.venv/bin/python M3_Train_AM.py --type DNN | tee train_dnn.log
../.venv/bin/python M3_Plot_Training.py --log train_dnn.log

For the full pipeline (batch features → LM/FST → decoding → scoring) see the step-by-step commands in ../Solutions/README.md.

Checking your answer

After finishing a script, diff it against the reference implementation:

diff M1_Score.py ../Solutions/M1_Score.py

Differences in the header comments are expected; focus on the logic inside the TODO(M#) blocks.

This site is open source. Improve this page.