seda

Bâki kalan bu kubbede bir hoş sadâ imiş...

— Bâkî

Lightweight, low-latency Turkish speech models for real-time applications.

Seda aims to provide models that are small enough to run on a CPU, fast enough for live audio, accurate enough for production, and open for commercial use. sedalib is the Python package for running them. It currently ships streaming speech recognition (STT) on top of sherpa-onnx.

Models

WER is measured with beam search (4 paths) and the language model enabled. The model processes audio in 0.64-second chunks, so partial results arrive in real time. See the model card for more benchmarks.

Installation

pip install sedalib

Python 3.10 or newer is required. The only runtime dependencies are sherpa-onnx and huggingface-hub.

Quick start

Transcribe a complete waveform. read_wav loads a 16-bit PCM WAV file as mono float samples in [-1, 1] together with its sample rate:

from sedalib import Recognizer, read_wav

samples, sample_rate = read_wav("audio.wav")

recognizer = Recognizer.from_pretrained()  # downloads the model once and caches it
text = recognizer.transcribe(samples, sample_rate)
print(text)

Audio at other sample rates is resampled automatically.

Streaming

For live audio, open a stream and feed it chunks as they arrive. accept returns the current partial transcript, and is_endpoint tells you when the speaker has finished an utterance.

stream = recognizer.stream()

for chunk in microphone_chunks():  # e.g. 0.64 s of audio at a time
    partial = stream.accept(chunk, sample_rate=16000)
    print(partial)

    if stream.is_endpoint:
        print("final:", stream.text)
        stream.reset()  # start a new utterance

Call stream.finish() at the end of the audio to flush the remaining frames and get the final text. A single Recognizer can serve many independent streams.

Options

Recognizer.from_pretrained(
    "atasoglu/seda-v0.1",  # Hugging Face repo id
    revision=None,  # pin a branch, tag or commit
    use_lm=True,  # rescore with the language model
    num_threads=1,  # CPU threads used for inference
)

The language model improves accuracy at the cost of some extra compute. Pass use_lm=False for the lightest setup. To load a model you have already downloaded, use Recognizer("path/to/model_dir").

Demo

A Gradio demo with live microphone streaming and file transcription is available as an extra:

pip install "sedalib[demo]"
python -m sedalib.demo.stt

Development

uv sync
uv run prek install  # ruff lint and format hooks

License

The code is released under Apache-2.0. The acoustic model is Apache-2.0 as well. The optional language model is CC BY-SA 4.0 because it was trained on Wikipedia data, so pass use_lm=False if that is a concern for your use case.