samplerate test about audfprint HOT 2 CLOSED

The numbers of matching and total hashes look extremely small. I wouldn't
trust any matches with fewer than 5-10 hash matches - 1 or 2 hashes are
very likely due to coincidences.

You should up your density until you're getting at least 20-40 total hashes
in your test segments. Also, it's not going to work well with very short
match patterns - I'd say 2 sec was a minimum. Single words sound too
little, and I don't really understand the use-case. Why would such a short
segment be repeated exactly?

It absolutely won't match different instances of the same word, even spoken
by the same speaker. That's a different recognition problem.

DAn.

On Wednesday, June 10, 2015, apiszcz [email protected] wrote:

I tested the samplerate for ingest of content and query values as follows.
Note Case 1 is correct. All source data is original 8KHZ.
--exact-count
--min-count 1
--density 100
--max-matches 10
--match-win 5
--pks-per-frame 2

Case1:--samplerate 11000, first case is the query itself in the content,
second instance is a match, 62.1 seconds is another instance, All
detections in case 1 are correct.

at 46.476 s with 2 of 9 hashes at rank 1
at 47.686 s with 1 of 9 hashes at rank 1
at 62.183 s with 1 of 9 hashes at rank 1

Case 2: Here we have a match with itself, however it misses matches at
47.6 and 62 in the top 10 results.

With 8KHZ data --samplerate 8000

at 46.048 s with 3 of 13 hashes at rank 1
at 332.416 s with 2 of 13 hashes at rank 1
at 12.672 s with 1 of 13 hashes at rank 1
at 13.472 s with 1 of 13 hashes at rank 1
at 97.408 s with 1 of 13 hashes at rank 1
at 121.056 s with 1 of 13 hashes at rank 1
at 146.304 s with 1 of 13 hashes at rank 1
at 257.216 s with 1 of 13 hashes at rank 1
at 323.008 s with 1 of 13 hashes at rank 1

—
Reply to this email directly or view it on GitHub
#9.

from audfprint.

Thank you for the feedback, use case is fingerprint detection on short
segments, I'm exploring parameter ranges for reasonable hit rates. Agree,
2-4 seconds seems more reasonable. Recognition, understand.

On Sat, Jun 13, 2015 at 1:20 AM, Dan Ellis [email protected] wrote:

The numbers of matching and total hashes look extremely small. I wouldn't
trust any matches with fewer than 5-10 hash matches - 1 or 2 hashes are
very likely due to coincidences.

You should up your density until you're getting at least 20-40 total hashes
in your test segments. Also, it's not going to work well with very short
match patterns - I'd say 2 sec was a minimum. Single words sound too
little, and I don't really understand the use-case. Why would such a short
segment be repeated exactly?

It absolutely won't match different instances of the same word, even spoken
by the same speaker. That's a different recognition problem.

DAn.

On Wednesday, June 10, 2015, apiszcz [email protected] wrote:

I tested the samplerate for ingest of content and query values as
follows.
Note Case 1 is correct. All source data is original 8KHZ.
--exact-count
--min-count 1
--density 100
--max-matches 10
--match-win 5
--pks-per-frame 2

Case1:--samplerate 11000, first case is the query itself in the content,
second instance is a match, 62.1 seconds is another instance, All
detections in case 1 are correct.

at 46.476 s with 2 of 9 hashes at rank 1
at 47.686 s with 1 of 9 hashes at rank 1
at 62.183 s with 1 of 9 hashes at rank 1

Case 2: Here we have a match with itself, however it misses matches at
47.6 and 62 in the top 10 results.

With 8KHZ data --samplerate 8000

at 46.048 s with 3 of 13 hashes at rank 1
at 332.416 s with 2 of 13 hashes at rank 1
at 12.672 s with 1 of 13 hashes at rank 1
at 13.472 s with 1 of 13 hashes at rank 1
at 97.408 s with 1 of 13 hashes at rank 1
at 121.056 s with 1 of 13 hashes at rank 1
at 146.304 s with 1 of 13 hashes at rank 1
at 257.216 s with 1 of 13 hashes at rank 1
at 323.008 s with 1 of 13 hashes at rank 1

—
Reply to this email directly or view it on GitHub
#9.

—
Reply to this email directly or view it on GitHub
#9 (comment).

from audfprint.