README.md

metadata

title: Omnivoice
emoji: 👁
colorFrom: pink
colorTo: blue
sdk: gradio
sdk_version: 6.22.0
python_version: '3.12'
app_file: app.py
pinned: false
short_description: omnivoice

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

Batch generation

One reference audio + one reference text + N target texts (one per line) → N generated clips, returned in input-line order.

Calling the API

The Space exposes a /batch_generate endpoint. Install the client with pip install gradio_client.

from gradio_client import Client, handle_file

# Space id, or "http://127.0.0.1:7860/" when running locally.
client = Client("your-username/omnivoice")

paths, status = client.predict(
    texts="The first line to generate.\n"
          "The second line to generate.\n"
          "And a third one.",
    ref_audio=handle_file("reference.wav"),   # or a URL
    ref_text="Transcript of the reference audio.",  # "" to auto-transcribe
    language="Auto",          # or e.g. "English"; applies to every line
    batch_size=8,             # lines per GPU call
    num_step=32,
    guidance_scale=2.0,
    denoise=True,
    speed=1.0,                # ignored when duration is set
    duration=None,            # fixed seconds per line; None = estimate
    preprocess_prompt=True,
    postprocess_output=True,
    api_name="/batch_generate",
)

print(status)
for i, path in enumerate(paths, start=1):
    print(i, path)   # local .wav downloaded by the client, in input order

paths is a list of downloaded .wav files — element i corresponds to line i of texts. On failure paths is empty and status carries the error message, so check it:

if not paths:
    raise RuntimeError(status)

Copy the files somewhere permanent; the client writes them to a temp directory:

import shutil

for i, path in enumerate(paths, start=1):
    shutil.copy(path, f"out_{i:03d}.wav")

Notes

• Only texts and ref_audio really matter; the rest have the defaults shown above. client.view_api() prints the exact signature if you want to confirm.
• Leaving ref_text="" runs ASR on the reference audio once, then reuses that transcript for every line. Passing the real transcript is faster and more accurate.
• Blank lines in texts are skipped. The maximum is 200 lines per call.
• On ZeroGPU each sub-batch is a separate GPU request, so a long list draws down your quota. Raise batch_size for fewer, larger requests; lower it if you hit out-of-memory errors.