euxoa
commented
3 years ago I stumbled into this only two days ago. With some data processing skills, you can import German names into the model/labels.txt file, if you have a list with matching scientific vs German names available. I did that for Finnish names, in R:
library(dplyr)
labels <- readr::read_delim("labels.txt", delim="_", col_names=c("name", "ename"))
transl <- readr::read_delim("../Maailman-lintujen-suomenkieliset-nimet-20180731.txt",
delim="\t", locale=readr::locale(encoding="latin1"))
labels %>%
left_join(transl %>% mutate(name=`Tieteellinen nimi`, fname=`Nimi suomeksi`)) %>%
select(name, fname, ename) %>%
mutate(cname=ifelse(is.na(fname), ename, fname)) %>%
mutate(line=paste(name, cname, sep="_")) %>%
{ paste(.$line, collapse="\n")} %>% {writeLines(., con="labels2.txt") }Just keep the format as name1_name2 for each line, and the positions of lines in the file intact.
What comes to reliability, I don't think you get complete reliability with any cutoff. I have used the default 0.1 and filtered out lines with reliability above 0.7 or 0.9 from the result file. Then if some of these are interesting, I look at the spectrogram in Audacity, and listen to it, to confirm or reject.
But the result file with cutoff of 0.1 is useful as well! For birds tend to have continuity, and typically the same bird appears in successive or almost successive frames. If you get something unexpected on the high-reliability list, look at the context around that frame in the file with cutoff=0.1. Often you find the correct id from the frames nearby (and it is a common bird).
Having been with this only a couple of nights, my workflow is definitely not final yet.
That said, I have already found interesting stuff from last year's records. Using BirdNET seems to have potential to speed up browsing of WAV files, and to reveal things that otherwise go unnoticed. Just don't take the ids at face value. They need to be confirmed somehow.
I think the identifications would be better if there were some continuity of scores over successive frames. It's not build into the current model, but maybe one can bring some of it there by post-processing. By looking at the source of analyze.py it would be best to take the whole score vector over several (say 3–10) frames and smooth it somehow, maybe by convolution in logit space. Or a Markov model. I just don't currently have a ground truth against which to optimise the required smoothness parameter. ;) A crude way is to manually look for identical labels in many subsequent frames. Have to think about this, although probably I wouldn't have time to code anything..
patlevin
nilspupils
ghost
DD4WH
I got BirdNET-Lite running and want to try to analyse a days worth of recording in my garden. The output generated is in latin and english. Do you have a german language labels.txt as in the phone app?
Also can you elaborate a little into the confidence values? Where would a "reliable" be ? at 0.5 or 0.9?
Thanks for the great work. I have so much fun "catching" new birds.