Quick Start

This page walks through a complete >> pipeline from loading data to viewing results.

Install

pip install plantain2asr[asr-cpu]

For GPU workflows: pip install plantain2asr[asr-gpu]

1. Load a dataset

from plantain2asr import GolosDataset

ds = GolosDataset("data/golos")
crowd = ds.filter(lambda s: s.meta["subset"] == "crowd")

2. Run models via `>>`

from plantain2asr import Models

crowd >> Models.GigaAM_v3()
crowd >> Models.Whisper()

Results are cached on disk -- reruns skip already processed samples.

3. Normalize via `>>`

from plantain2asr import SimpleNormalizer

norm = crowd >> SimpleNormalizer()

The original dataset is untouched; norm is a new view with normalized texts.

4. Compute metrics via `>>`

from plantain2asr import Metrics

norm >> Metrics.composite()

This computes WER, CER, MER, WIL, WIP, Accuracy, IDR, and LengthRatio in a single pass.

5. Explore results

PandasInteractive reportStatic HTML reportCSV export

df = norm.to_pandas()
print(df.groupby("model")[["WER", "CER", "Accuracy"]].mean().sort_values("WER"))

from plantain2asr import ReportServer
ReportServer(norm, audio_dir="data/golos").serve()

from plantain2asr.reporting.builder import ReportBuilder
ReportBuilder(norm).save_static_html("artifacts/report.html")

norm.save_csv("artifacts/results.csv")

Full pipeline in one block

from plantain2asr import GolosDataset, Models, SimpleNormalizer, Metrics, ReportServer

ds = GolosDataset("data/golos")

ds >> Models.GigaAM_v3()
ds >> Models.Whisper()

norm = ds >> SimpleNormalizer()
norm >> Metrics.composite()

df = norm.to_pandas()
print(df.groupby("model")[["WER", "CER"]].mean().sort_values("WER"))

ReportServer(norm, audio_dir="data/golos").serve()

Loading precomputed results

If inference was run on another machine, load JSONL results and continue evaluation locally:

ds = GolosDataset("data/golos")
ds.load_model_results("GigaAM-v3-rnnt", "results/GigaAM-v3-rnnt_results.jsonl")

One line per sample:

{"audio_path": "/any/path/file.wav", "hypothesis": "recognized text", "processing_time": 1.23}

Samples are matched by the basename of audio_path, which makes cross-machine reuse practical.

`Experiment` convenience wrapper

If you want ready-made research scenarios without building the >> chain manually, Experiment wraps the same pipeline steps:

from plantain2asr import Experiment, GolosDataset, Models, SimpleNormalizer

experiment = Experiment(
    dataset=GolosDataset("data/golos"),
    models=[Models.GigaAM_v3(), Models.Whisper()],
    normalizer=SimpleNormalizer(),
)

experiment.compare_on_corpus(metrics=["WER", "CER", "Accuracy"])

Method	What it does
`compare_on_corpus()`	Run models, normalize, score, return comparison table
`leaderboard()`	Rank models by a single metric
`prepare_thesis_tables()`	Export CSV tables for thesis/paper
`export_appendix_bundle()`	Full package: tables + report + benchmark
`benchmark_models()`	Latency, throughput, RTF measurements
`save_report_html()`	Static HTML report

Under the hood, Experiment executes the same >> steps. Use it when you want a one-liner; use the pipeline when you want control.