Skip to content

plantain2asr

Benchmarking and analysis framework for Russian ASR models.

plantain2asr is built around the >> pipeline operator: you load a dataset, push it through models, normalizers, and metrics -- each step produces a new view, nothing is mutated in place.

The >> interface

from plantain2asr import GolosDataset, Models, SimpleNormalizer, Metrics

ds = GolosDataset("data/golos")

ds >> Models.GigaAM_v3()      # run inference, results are cached
ds >> Models.Whisper()         # run another model on the same data

norm = ds >> SimpleNormalizer()  # normalize references and hypotheses
norm >> Metrics.composite()      # compute WER, CER, MER, Accuracy, ...

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

Each >> step returns the dataset with new results layered on top. You can branch, filter, and recombine at any point.

What plantain2asr gives you

  • The >> pipeline: dataset >> model >> normalizer >> metric >> report
  • Local and cloud ASR backends under one interface
  • Automatic device choice (CUDA / MPS / CPU) where supported
  • Immutable dataset views instead of in-place mutation
  • Built-in normalization, metrics, reports, analysis, and benchmarks
  • Experiment convenience wrapper for common research scenarios
  • Modular architecture for adding your own models, metrics, and report sections

Install

pip install plantain2asr
Includes datasets, normalization, metrics, exports, and reporting.

pip install plantain2asr[asr-cpu]

pip install plantain2asr[asr-gpu]

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"

pip install plantain2asr[analysis]

pip install plantain2asr[all]

Device resolution prefers CUDA, then MPS, then CPU where the backend supports it.

Mental Model

graph LR
    A["ds = GolosDataset(...)"] -->|>>| B["Models.GigaAM_v3()"]
    A -->|>>| B2["Models.Whisper()"]
    B --> C[">> SimpleNormalizer()"]
    B2 --> C
    C -->|>>| D["Metrics.composite()"]
    D --> E["to_pandas() / ReportServer / CSV"]

Every block is a >> step. Compose them however you need.

Supported model families

Family Typical call Extra Device
GigaAM v3 Models.GigaAM_v3() gigaam CUDA / MPS / CPU
GigaAM v2 Models.GigaAM_v2() gigaam CUDA / MPS / CPU
Whisper Models.Whisper() whisper CUDA / MPS / CPU
T-one Models.Tone() tone + T-One source archive CUDA / CPU
Vosk Models.Vosk(...) vosk CPU
Canary Models.Canary() canary CUDA
SaluteSpeech Models.SaluteSpeech() none cloud

If you are new