plantain2asr 1.0.4

A benchmarking and analysis framework for Russian ASR models

🌱 plantain2asr

Benchmarking and analysis framework for Russian ASR models.

plantain2asr lets you compare ASR backends, normalize transcripts, compute metrics, inspect errors, benchmark latency, and export research artifacts without losing the underlying composable pipeline model.

Start Here

There are three entry points, ordered from simplest to most flexible:

1. Interactive Constructor: open the docs constructor and assemble a ready-made chain visually.
2. Experiment facade: run common research scenarios with a few high-level calls.
3. >> pipeline API: build custom chains from datasets, models, normalizers, metrics, reports, and analyzers.

Docs: akatsnelson.github.io/plantain2asr

Install

# Core only: datasets, normalization, metrics, reports
pip install plantain2asr

# Common CPU-only local stack
pip install plantain2asr[asr-cpu]

# Common GPU-ready local stack
pip install plantain2asr[asr-gpu]

# Individual model families
pip install plantain2asr[gigaam]
pip install plantain2asr[whisper]
pip install plantain2asr[vosk]
pip install plantain2asr[canary]
pip install plantain2asr[tone]
pip install "tone @ https://github.com/voicekit-team/T-one/archive/3c5b6c015038173840e62cea99e10cdb1c759116.tar.gz"

# Research analysis tools
pip install plantain2asr[analysis]

# Everything
pip install plantain2asr[all]

Device selection is automatic where supported: NVIDIA GPU first, then MPS, then CPU.

Recommended Quick Start

For most research workflows, start with the >> pipeline:

from plantain2asr import GolosDataset, Models, SimpleNormalizer, Metrics

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", "Accuracy"]].mean().sort_values("WER"))

If you want ready-made research scenarios on top of the same building blocks, use Experiment:

• Experiment.compare_on_corpus() for straightforward model comparison
• Experiment.prepare_thesis_tables() for publication-ready aggregate tables
• Experiment.export_appendix_bundle() for a full appendix bundle with exports and optional static report
• Experiment.benchmark_models() for latency, throughput, and RTF measurement

Advanced Pipeline API

If you want full composability, the canonical chain is still:

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

ds = GolosDataset("data/golos")
ds >> Models.GigaAM_v3()
ds >> Models.Whisper()

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

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

Pipeline rules:

• Every step returns a dataset or processor-compatible object.
• Normalization creates a new dataset view and does not mutate the original.
• Model results are cached and safe to resume.
• You can branch at any point with filter(), take(), or cloned views.

Supported Models

Call	Stored name	Extra	Device
Models.GigaAM_v3()	GigaAM-v3-e2e_rnnt	gigaam	CUDA / MPS / CPU
Models.GigaAM_v3(model_name="e2e_ctc")	GigaAM-v3-e2e_ctc	gigaam	CUDA / MPS / CPU
Models.GigaAM_v3(model_name="rnnt")	GigaAM-v3-rnnt	gigaam	CUDA / MPS / CPU
Models.GigaAM_v3(model_name="ctc")	GigaAM-v3-ctc	gigaam	CUDA / MPS / CPU
Models.GigaAM_v2(model_name="v2_rnnt")	GigaAM-v2_rnnt	gigaam	CUDA / MPS / CPU
Models.GigaAM_v2(model_name="v2_ctc")	GigaAM-v2_ctc	gigaam	CUDA / MPS / CPU
Models.Whisper()	Whisper-whisper-large-v3-ru-podlodka	whisper	CUDA / MPS / CPU
Models.Tone()	T-One	tone + T-One source archive	CUDA / CPU
Models.Vosk(model_path=...)	Vosk	vosk	CPU
Models.Canary()	Canary-1B	canary	CUDA
Models.SaluteSpeech()	SaluteSpeech	none	cloud

You can also resolve models by user-facing names with Models.create(...), including case and separator variants such as "gigaam_v3", "GigaAM-v3", or "tone".

Typical Research Outputs

• Metrics tables as Python dicts or pandas DataFrames
• Leaderboards sorted by a chosen metric
• Error-case tables and CSV exports
• Static HTML reports for sharing without a running server
• Appendix bundles with thesis-ready artifacts
• Benchmark summaries for CPU, CUDA, or MPS

Extending

plantain2asr keeps the "plantain" idea of modular composition. If the built-in stack is not enough, extend one of the four base types:

• BaseASRModel
• BaseNormalizer
• BaseMetric
• BaseSection

See the docs extending guides for custom components and implementation patterns.

License

MIT — Artem Katsnelson