fasr-vad-firered 0.5.2

FireRedVAD for fasr (bundled fireredvad inference)

fasr-vad-firered

Chinese documentation

FireRedVAD voice activity detection for fasr. This is an offline neural VAD that loads FireRed's PyTorch checkpoint and returns AudioSpan speech segments.

Install

pip install fasr-vad-firered

Registered Model

Registry name	Class	Best for
firered	FireRedForVAD	Offline VAD with FireRed checkpoints

The default checkpoint is FireRedTeam/FireRedVAD. Local checkpoint directories must contain the upstream VAD files, typically cmvn.ark and model.pth.tar or a VAD/ subdirectory containing them.

Pipeline Usage

from fasr import AudioPipeline

pipeline = (
    AudioPipeline()
    .add_pipe(
        "detector",
        model="firered",
        speech_threshold=0.4,
        use_gpu=False,
    )
    .add_pipe("recognizer", model="firered_aed")
    .add_pipe("sentencizer", model="ct_transformer")
)

Quick choices:

Goal	Use	Result
Reduce noise false positives	speech_threshold=0.55	Requires stronger speech posterior
Keep weak speech	speech_threshold=0.3	More sensitive, but may include noise
Use GPU inference	use_gpu=True	Faster when CUDA is available

Confection Config

[vad_model]
@vad_models = "firered"
use_gpu = false
speech_threshold = 0.4

Inside a pipeline:

[pipeline]
@pipelines = "AudioPipeline.v1"
pipe_order = ["detector"]

[pipeline.pipes]

[pipeline.pipes.detector]
@pipes = "thread_pipe"

[pipeline.pipes.detector.component]
@components = "detector"

[pipeline.pipes.detector.component.model]
@vad_models = "firered"
use_gpu = false
speech_threshold = 0.4

Direct Model Usage

from fasr.config import registry
from fasr.data import AudioSpan, Waveform

model = registry.vad_models.get("firered")(
    speech_threshold=0.4,
    use_gpu=True,
)

audio = AudioSpan(waveform=Waveform.from_file("example.wav"), start_ms=0)
segments = model.detect(audio)
for segment in segments:
    print(f"{segment.start_ms}ms - {segment.end_ms}ms")

Use local weights:

model.load_checkpoint("/path/to/FireRedVAD")

Parameters

Parameter	Type / range	Default	Higher value	Lower value	Change when
use_gpu	bool	False	Enables CUDA inference	Uses CPU	You have CUDA available and need speed
speech_threshold	float, 0.0 to 1.0	0.4	More conservative; fewer false positives	More sensitive; more weak speech retained	Noise leaks in, or speech is missed

Generic checkpoint fields such as checkpoint, cache_dir, endpoint, revision, and force_download are inherited from the base model.

Tuning Guide

Symptom	Try first
Noise is detected as speech	Raise speech_threshold to 0.5 or 0.6
Quiet speech is missed	Lower speech_threshold to 0.3
CPU inference is too slow	Set use_gpu=True on a CUDA machine

Dependencies

• fasr
• torch >= 2.0.0
• soundfile >= 0.12.0
• kaldiio >= 2.18.0
• kaldi-native-fbank >= 1.19.0
• Python 3.10-3.12