Shell‑script vs Python workflow in Kaldi: trade‑off between transparency and newcomer friendliness

Recommendation

Adopt the native shell‑script workflow as the default user‑facing entry point. It preserves full command‑line transparency and exact reproducibility, which are core strengths of Kaldi. Provide the Python wrapper as an optional, higher‑level alternative for newcomers who prefer NumPy/PyTorch tensors, but ship it with a logging mode that exposes the underlying shell commands.

Steps to Standardize the Python Wrapper Across OSes

1. Pin dependencies: maintain a requirements.txt (or environment.yml for conda) that lists exact versions of kaldi_io, kaldialign, numpy, and torch tested against a specific Kaldi release (e.g., Kaldi 5.5).
2. Use a locked environment: instruct users to create a fresh environment from the locked file before running any wrapper scripts.
3. Test on target platforms: run the wrapper’s smoke‑test on a clean Ubuntu, macOS, and Windows VM with the same Python and Kaldi versions; record any failures and adjust pins accordingly.
4. Enable command logging: add an environment variable (e.g., KALDI_LOG_CMDS=1) or a flag (--log-commands) to the wrapper that, before each Kaldi binary call, prints the full subprocess.run([...]) argument list to stdout or a log file. This makes the wrapper’s actions auditable without sacrificing the high‑level API.

Verification of Reproducibility

To confirm that the wrapper behaves identically to the shell‑script path:

• Run a standard feature extraction (e.g., steps/make_mfcc.sh) using the native script and record the output .ark files.
• Run the equivalent extraction via the Python wrapper (with logging enabled) and compare the resulting .ark files using compute‑wer.pl --mode ark or a simple diff on the binary matrices.
• If the outputs match bit‑for‑bit, the wrapper’s abstraction is verified for that Kaldi version.

Missing Diagnostic Detail

The recommendation assumes a specific Kaldi release and target operating system. Please specify the Kaldi version (e.g., 5.5, master) and the OS distribution you intend to support so the dependency pins and compatibility tests can be tailored accordingly.