Internet radio speech music discrimination using libopus and ffmpeg.
This script wraps ffmpeg and the analysis executable and exports the results as Prometheus metrics.
[icecast stream] -> [ffmpeg] -> [opus_sm] -> [prometheus]
^^^^^^^^^^^^^^^^^^^^^^^ (sm_exporter scope)
You'll need a compiled opus_sm_demo executable and ffmpeg installed. This script forks them using submodule.
P("Content is music")
# HELP sm_music_probability Probability of music in stream
# TYPE sm_music_probability gauge
sm_music_probability{url="..."} 0.999795
This is a histogram measuring time between analysis results, you can use it to select an appropriate scrape_interval.
# HELP sm_analysis_result_s Instrumenting analysis executable output
# TYPE sm_analysis_result_s histogram
sm_analysis_s_bucket{le="0.005",url="..."} 26.0
...
sm_analysis_result_s_bucket{le="+Inf",url="..."} 791.0
sm_analysis_result_s_count{url="..."} 791.0
sm_analysis_result_s_sum{url="..."} 156.07464370007
This counter is incremented when ffmpeg or opus_sm_demo exit, resulting in them being reinitialized.
# HELP sm_stream_retries_total Incremented on retry
# TYPE sm_stream_retries_total counter
sm_stream_retries_total{url="..."} 0.0
You can find a fork of opus_sm that adds support for streaming here:
https://github.com/aleksipirttimaa/opus_sm
The build instructions in the aforementioned repository work just fine. On Debian you'll need to install some packages:
sudp apt install build-essential automake libtoolJust to recap building:
./autogen.sh
./configure
makethe resulting binary is ./opus_sm_demo
- Make sure you have
opus_sm_demoandffmpeg
which opus_sm_demo && which ffmpeg- Install pipenv, which is used for dependency management
python -m pip install --user pipenv- Create virtualenv and install dependencies
python -m pipenv install- run src/main.py
python -m pipenv run python src/main.py https://stream.example.com/example.m3u8Add to your prometheus.yml:
scrape_configs:
- job_name: 'speechmusic'
scrape_interval: 1s
static_configs:
- targets: ['localhost:9069']
A good scrape_interval depends on the incoming stream transport and encoding, as analysis can only be performed after an entire frame of the encoded audio has been received.