Performance

Status: Current Last updated: 2026-07-30 18:21 EDT

This page covers what to expect from Batchalign’s processing times and how to improve throughput.

Cold vs warm starts

The first run of any command downloads ML models and initializes them, expect 5-20x longer than subsequent runs. After the first run:

• Model cache: Stanza, Whisper, and other ML models are cached on disk (~2 GB total). They load from cache on subsequent runs.
• Server warmth: When an explicit server is running, workers can stay warm in memory across multiple jobs. Direct local execution does not keep a daemon alive between CLI invocations.
• Analysis cache: Batchalign caches audio-bound intermediate results (forced-alignment word timings, UTR ASR) in a local SQLite database keyed by content hash. Re-running align or transcribe on the same audio reuses these and is much faster. Text-NLP commands (morphotag, utseg, translate, coref) are not cached: see Caching.

Worker count

By default, Batchalign uses one worker per command. For batch processing of many files, increase the worker count:

batchalign3 morphotag ~/corpus/ -o ~/output/ --workers 4

Each worker loads its own copy of the ML models. Memory usage scales linearly with worker count, see the memory section below.

CPU vs GPU

Batchalign automatically uses GPU acceleration when available (CUDA on Linux, MPS on macOS). To force CPU-only processing:

batchalign3 morphotag ~/corpus/ -o ~/output/ --force-cpu

CPU-only is slower but uses less memory and avoids GPU driver issues. On machines without a supported GPU, CPU mode is selected automatically.

Memory patterns

Memory usage depends on the command and number of workers:

With --workers N, total memory is roughly N * per-worker cost. The Rust runtime adds minimal overhead (~50 MB).

Lazy audio loading: Audio files are loaded on demand and released after processing, memory does not grow with corpus size, only with concurrent workers.

Server mode for warm models

For repeated interactive use, keep models loaded in the background:

batchalign3 serve start

Subsequent commands automatically connect to the running daemon. Stop it when done:

batchalign3 serve stop

See Server Mode for configuration details and Worker Tuning for memory budgets and tuning.

The bench command

Measure processing throughput on your hardware. The shape is bench <command> <in_dir> <out_dir>: both directories are required positional arguments:

batchalign3 bench morphotag ~/sample-corpus/ ~/bench-out/ --workers 1
batchalign3 bench morphotag ~/sample-corpus/ ~/bench-out/ --workers 4

This runs the command with timing instrumentation and reports files/second and wall-clock time per file. Use --runs N to repeat the run, --use-cache to keep cache lookups enabled (the default is to bypass cache for clean benchmarks), and --dataset <label> to tag structured output.

Estimated times per command

Rough estimates for a single file (~100 utterances) on a modern laptop with warm daemon:

Times vary significantly with hardware, file size, and language. GPU acceleration typically provides a 2-5x speedup for model inference.

This page last changed: 2026-07-31 (commit 8e3b0e23). The whole book last changed: 2026-09-16 (commit 34d249d8).