FireRedVAD

FireRedVAD

Detects speech, singing and music in audio, with streaming and file-based voice activity detection plus audio event detection. For speech developers who segment or clean audio before transcription.

Made by FireRedTeam on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: FireRedVAD finds where speech happens in an audio file, and it can also flag singing and music. It offers voice activity detection in streaming and non-streaming modes, plus non-streaming audio event detection, and the card says it covers more than 100 languages. It was introduced as part of FireRedASR2S and is also released as a standalone project with weights and inference code.

Who it suits: Speech and audio developers who need to cut recordings into speech segments or tell speech from singing and music.

Industry | Who uses it | How they use it | Example
Speech technology | ASR developers | Cut audio into speech segments before transcribing | Drop the silence from a recorded call
Voice products | Product teams | Detect speech as audio streams in | Start listening only when someone talks
Media | Audio editors | Find speech, singing and music in a track | Mark where singing begins
Research | Speech researchers | Compare detectors on a multilingual benchmark | Check false alarms against Silero-VAD

You can skip it if you need live audio event detection, since that part works on whole files only.

Where to use it: It runs from the command line with Python, and the card gives example shell scripts for each mode.

• A command line with conda and Python 3.10
• Windows, macOS or Linux with Python, to download the model files
• ffmpeg, to convert audio to 16 kHz mono when needed

Use case | Who | What they get back
File voice detection | ASR developers | Speech segments written to a text file and clip folder
Streaming detection | Voice product teams | Speech segments found as audio arrives
Audio event detection | Media teams | Labels for speech, singing and music

How to set it up: Make a clean Python environment, get the code project, install its requirements, then download the models. Audio should be 16 kHz, 16-bit, mono PCM.

• conda and Python 3.10
• The FireRedVAD code project from GitHub
• ffmpeg, if your audio needs converting

1. Create a clean conda environment with Python 3.10 and activate it.
2. Get the FireRedVAD code from GitHub and open its folder.
3. Install the requirements and set PATH and PYTHONPATH.
4. Download the models from ModelScope or Hugging Face.
5. Convert your audio to 16 kHz 16-bit mono PCM if needed.

conda create --name fireredvad python=3.10
conda activate fireredvad

pip install -r requirements.txt
export PATH=$PWD/fireredvad/bin/:$PATH
export PYTHONPATH=$PWD/:$PYTHONPATH

pip install -U modelscope
modelscope download --model xukaituo/FireRedVAD --local_dir ./pretrained_models/FireRedVAD

pip install -U "huggingface_hub[cli]"
huggingface-cli download FireRedTeam/FireRedVAD --local-dir ./pretrained_models/FireRedVAD

ffmpeg -i <input_audio_path> -ar 16000 -ac 1 -acodec pcm_s16le -f wav <output_wav_path>

How to use it: Run the example scripts from the examples folder, or call the command-line tools yourself. Each tool has a help option that lists its flags.

cd examples
bash inference_vad.sh
bash inference_stream_vad.sh
bash inference_aed.sh

vad.py --help
vad.py --use_gpu 0 --model_dir pretrained_models/FireRedVAD/VAD --smooth_window_size 5 --speech_threshold 0.4 \
    --min_speech_frame 20 --max_speech_frame 3000 --min_silence_frame 10 --merge_silence_frame 0 \
    --extend_speech_frame 0 --chunk_max_frame 30000 --write_textgrid 1 \
    --wav_path assets/hello_zh.wav --output out/vad.txt --save_segment_dir out/vad

Tool | What it does
vad.py | Voice activity detection on whole audio files
stream_vad.py | Streaming voice activity detection
aed.py | Audio event detection for speech, singing and music
speech_threshold | The cutoff that decides whether a frame counts as speech
write_textgrid | Writes the result as a TextGrid when set to 1

Example: Run vad.py on the sample file assets/hello_zh.wav with the card's settings. It writes the speech segments to out/vad.txt and saves each segment as a clip in out/vad.

For newcomers and experts: Voice activity detection answers a simple question, which parts of this recording contain someone talking. A first small try is to run the example script for file detection on the sample audio and read the output file.

For experts, the models are built on DFSMN and come as non-streaming VAD, streaming VAD and non-streaming AED. The card evaluates them on FLEURS-VAD-102, its own benchmark of 9,443 annotated files in 102 languages. FireRedVAD trades a slightly higher miss rate than TEN-VAD for a much lower false alarm rate than FunASR-VAD.

Convert audio to 16 kHz mono PCM
Pick VAD, streaming VAD or AED
Run the tool with its thresholds
Read segments or event labels

Good to know: The card compares F1 scores on FLEURS-VAD-102 for non-streaming voice activity detection.

Title: F1 score on FLEURS-VAD-102
FireRedVAD: 97.57%
Silero-VAD: 95.95%
TEN-VAD: 95.19%
FunASR-VAD: 90.91%
WebRTC-VAD: 52.30%

• Own benchmark: FLEURS-VAD-102 was built by the card's authors, so results on your audio may differ.
• Miss rate: 3.62%, against 2.95% for TEN-VAD and 0.42% for FunASR-VAD.
• False alarms: 2.69%, against 44.03% for FunASR-VAD.
• Audio format: 16 kHz, 16-bit, mono PCM.
• License: Apache-2.0.
• Cost to run: none named, no API key, account or paid service.
• Data it reaches: the audio files you point it at, and it writes results to local folders.
• Not for: live audio event detection, since AED works on whole files.

On Hugging Face, 13 November 2026

• Task: voice activity detection
• Downloads in the last 30 days: 2,286
• Likes: 80
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 8
• Fine-tunes: 4
• Tags: voice-activity-detection, speech-activity-detection, audio-event-detection, streaming, non-streaming, audio, automatic-speech-recognition

Links

• Model: https://huggingface.co/FireRedTeam/FireRedVAD
• Files: https://huggingface.co/FireRedTeam/FireRedVAD/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download FireRedTeam/FireRedVAD

Then follow the usage in its model card. Nothing in this kit runs by itself.