Description #

Speech endpoint detection module – an efficient audio segmentation solution based on Silero VAD

Key features #

• Speech endpoint detection: Automatically identifies the start and end times of speech segments in audio, enabling precise segmentation of speech and silence.
• Batch processing: Supports efficient endpoint detection for large volumes of audio data.
• Multi-format input: Compatible with various audio input methods, including raw binary, Base64 encoding, and TOS/HTTP links.
• GPU acceleration: Supports high-performance inference in GPU environments.
• High versatility: Delivers excellent performance when processing audio from different domains, with various background noise and quality levels.

Recommended practices #

• It is recommended to use WAV format audio with an 8k or 16k sampling rate and mono channel to improve detection accuracy.
• GPU acceleration does not provide a significant speedup; in scenarios where extreme performance is not required, you can use CPU inference with the onnx model.
• Applicable to scenarios such as voice activity detection and speech segmentation.

Supported models #

• silero_vad.onnx (corresponding to onnx_model_revision=16)
• silero_vad_16k_op15.onnx (corresponding to onnx_model_revision=15)
• silero_vad.jit

Output description #

• For each audio, a two-dimensional list of floating-point numbers is output, representing the start and end timestamps (in seconds) of all speech segments, for example: [[0.0, 4.34], [5.50, 7.12]].
• Returns None if processing fails.

Operator parameters #

Input #

Output #

A column containing speech endpoint timestamps, where each element is a list of two floating-point numbers representing the start and end timestamps (in seconds) of each speech segment in the audio,
for example: [[0.0, 4.34], [5.50, 7.12]]. Returns None if processing fails.

Parameters #

If a parameter does not have a default value, it is required.

Examples #

The following code demonstrates how to use daft to run the operator to identify voice endpoints in audio.

from __future__ import annotations

import logging
import os

import ray

import daft
from daft import col
from daft.las.functions.audio.audio_vad_silero import AudioVadSilero
from daft.las.functions.udf import las_udf

if __name__ == "__main__":
    if os.getenv("DAFT_RUNNER", "ray") == "ray":

        def configure_logging():
            logging.basicConfig(
                level=logging.INFO,
                format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
                datefmt="%Y-%m-%d %H:%M:%S.%s".format(),
            )
            logging.getLogger("tracing.span").setLevel(logging.WARNING)
            logging.getLogger("daft_io.stats").setLevel(logging.WARNING)
            logging.getLogger("DaftStatisticsManager").setLevel(logging.WARNING)
            logging.getLogger("DaftFlotillaScheduler").setLevel(logging.WARNING)
            logging.getLogger("DaftFlotillaDispatcher").setLevel(logging.WARNING)

        import ray

        ray.init(dashboard_host="0.0.0.0", runtime_env={"worker_process_setup_hook": configure_logging})
        daft.set_runner_ray()

    daft.set_execution_config(actor_udf_ready_timeout=600)
    daft.set_execution_config(min_cpu_per_task=0)

    tos_dir_url = os.getenv("TOS_DIR_URL", "las-cn-beijing-public-online.tos-cn-beijing.volces.com")
    samples = {
        "audio_path": [
            f"https://{tos_dir_url}/public/shared_audio_dataset/.wav"
        ]
    }

    model_path = os.getenv("MODEL_PATH", "/opt/las/models")
    model_name = "silero-vad"
    audio_src_type = "audio_url"
    use_onnx_model = True
    onnx_model_revision = 16

    df = daft.from_pydict(samples)
    df = df.with_column(
        "audio_vad_result",
        las_udf(
            AudioVadSilero,
            construct_args={
                "audio_src_type": audio_src_type,
                "model_path": model_path,
                "model_name": model_name,
                "use_onnx_model": use_onnx_model,
                "onnx_model_revision": onnx_model_revision,
            },
            num_gpus=1,
            batch_size=1,
            concurrency=1,
        )(col("audio_path")),
    )
    df.show()

    # ╭────────────────────────────────┬─────────────────────╮
    # │ audio_path                     ┆ audio_vad_result    │
    # │ ---                            ┆ ---                 │
    # │ Utf8                           ┆ List[List[Float32]] │
    # ╞════════════════════════════════╪═════════════════════╡
    # │tos://las-cn-beijing-public-on… ┆ [[0.8, 2.4]]        │
    # ╰────────────────────────────────┴─────────────────────╯