Skip to content

README.md

Latest commit

 

History

History
83 lines (60 loc) · 1.93 KB

README.md

File metadata and controls

83 lines (60 loc) · 1.93 KB
Raw

Seda for Python

Typed Python 3.11+ client for a locally running Seda sidecar.

Install

From a checkout:

python -m pip install ./sdks/python

From GitHub:

python -m pip install \
  "bearlyai-seda @ git+https://github.com/bearlyai/seda.git#subdirectory=sdks/python"

Prepare and start the native helper once:

seda prepare \
  --model-id nvidia/nemotron-3.5-asr-streaming-0.6b \
  --variant q4_k
seda serve \
  --model-id nvidia/nemotron-3.5-asr-streaming-0.6b \
  --variant q4_k

Read the first JSON stdout line and pass its address and token to the client. Do not hard-code or persist that ephemeral token.

Stream application-owned audio

Prepare and start Seda once, then choose the language for each transcription or live stream:

from seda import Seda

seda = Seda.connect("http://127.0.0.1:7331", token="...")
print(seda.capabilities().resolved_model.id)

with seda.listen(language="de-DE") as session:
    session.write(pcm_s16le)
    transcript = session.commit(
        on_transcript=lambda update: print(update.text, end="\r")
    )

print(transcript.text)

write() accepts raw 16 kHz, mono, signed 16-bit little-endian PCM bytes. Transcript callbacks are revisions; render update.text as a replacement rather than appending it.

Transcribe a WAV file

For complete WAV files:

with open("speech.wav", "rb") as audio:
    transcript = seda.transcribe(audio.read(), language="en-US")

Build a language picker

capabilities = seda.capabilities()
print(
    capabilities.resolved_model.id,
    capabilities.resolved_model.revision,
    capabilities.resolved_model.variant,
)

if capabilities.language.mode == "prompted":
    choices = capabilities.language.supported

The model is selected when the Seda process is prepared and started. Language is selected independently for every call or stream. SedaError exposes stable code, recoverable, and HTTP status fields.