eist

Scottish Gaelic speech-to-subtitles pipeline — transcribe, align, punctuate, translate.

eist (Scottish Gaelic for "listen") is a Python library that packages all features of the ÈIST demo into a pip-installable local package following PyTorch / Hugging Face conventions.

Documentation · Models · Demo

Quick start

pip install eist[all]

End-to-end pipeline

from eist import GaelicASRPipeline

pipe = GaelicASRPipeline.from_pretrained()
result = pipe("audio.wav")

for sub in result:
    print(f"[{sub.start:.1f}–{sub.end:.1f}] {sub.text}")

# Export as SRT
with open("output.srt", "w") as f:
    f.write(result.to_srt())

Component-level usage

from eist import Transcriber, PunctuationModel, SubtitleFormatter
from sk_align import Aligner

# Load individual components
transcriber = Transcriber.from_pretrained()
aligner = Aligner.from_pretrained()
punct = PunctuationModel.from_pretrained()
formatter = SubtitleFormatter(max_words=10, min_words=3)

# Run step-by-step
segments = transcriber.transcribe("audio.wav")
# ... align, punctuate, format as needed

Translation

from eist import Translator

translator = Translator(api_key="sk-...")  # or set OPENAI_KEY env var
translator.translate(result.subtitles)

# Export translated SRT
print(result.to_srt(translated=True))

Installation

Install only the components you need:

pip install eist                    # core types + subtitle formatter (no heavy deps)
pip install eist[transcribe]        # + Whisper (faster-whisper, torch)
pip install eist[align]             # + forced alignment (sk-align)
pip install eist[punctuate]         # + punctuation model (onnxruntime)
pip install eist[all]               # everything

Command line

After installing, the eist command is available:

# Transcribe to SRT (default)
eist transcribe audio.wav -o output.srt

# WebVTT format
eist transcribe audio.wav -f vtt -o output.vtt

# Plain text transcript
eist transcribe audio.wav -f txt

# JSON output (includes word-level timestamps)
eist transcribe audio.wav -f json -o output.json

# Skip alignment or punctuation for faster processing
eist transcribe audio.wav --no-align --no-punctuate

# Translate to English (requires OpenAI API key)
eist transcribe audio.wav --translate --api-key sk-...

# Use a specific device / compute type
eist transcribe audio.wav --device cpu --compute-type float32

# Suppress progress messages
eist transcribe audio.wav -q -o output.srt

Run eist transcribe --help for all options. You can also use python -m eist.

Components

Data types

All components use typed dataclasses:

• Word — text, start, end
• Subtitle — text, start, end, words, paragraph_break, translation
• TranscriptionResult — subtitles, language, duration, timing
  • .to_srt(), .to_vtt() for export
  • .text for plain-text transcript
  • Iterable: for sub in result: ...

Models

All models are hosted on the eist-edinburgh Hugging Face organisation.

ASR (Speech Recognition)

*WER measured on 30 s of a BBC Radio nan Gàidheal broadcast, compared against a word-count-matched manual transcript. RTF = real-time factor (lower is faster); measured on CPU (Intel® Core™ i7, single-thread CTranslate2).

Forced Alignment

Alignment uses sk-align with a Kaldi nnet3 chain model.

Punctuation Restoration

Punctuation F1 scores averaged over two ~30 min BBC Radio nan Gàidheal broadcasts (~9 300 words total).

Benchmark details

All benchmarks run on an Intel® Core™ i7 CPU with float32 precision, no GPU. Audio from BBC Radio nan Gàidheal broadcasts, ~30 min each. ASR evaluation capped to 30 s due to CPU speed; alignment capped to 60 s. Full results in benchmarks/.

License

MIT