readbeyond
commented
6 years ago Hi,
the x86 limitation is known and due to maximum contiguous allocation size, and I consider it a "won't fix" for aeneas v1.x. The next major, aeneas v2.x will have external memory (i.e., on-disk) support, addressing the issue on x86.
The other problem on x64 under Windows (10? 8? 7?) is definitely more alarming. At first, I thought of an issue with eSpeak, since the number of frames extracted is about 2/3 of what it should --- but since you say that v1.7.0 works fine, I doubt my guess. A second hypothesis is a regression in the MFCC C code, but on top of my head I do not see a reason for it, and that code is pretty much OS-agnostic.
Luckily I have some time during the next days AND a Window 7 x64 laptop with me to investigate the problem. Hopefully I will have at least a diagnosis soon, thank you for reporting,
Alberto Pettarin
On 12/28/2017 08:32 PM, Nickolay wrote:
More precisely, Windows x64 and Python x64. It works fine with Python x86 but audio duration limit is much, much shorter (maybe ~30 minutes) in that case.
Maybe there's something wrong with DTWAligner because I noticed these lines in the log file:
|[DEBU] DTWAligner: Number of MFCC frames in real wave: 95000 [DEBU] DTWAligner: Number of MFCC frames in synt wave: 66469 |
For comprasion, these lines from the exact same file but on Ubuntu when I got correct timings.
|[DEBU] DTWAligner: Number of MFCC frames in real wave: 95000 [DEBU] DTWAligner: Number of MFCC frames in synt wave: 92897 |
How to produce:
|mkdir output set AENEAS_WITH_CEW=False pip install aeneas python -m aeneas.tools.execute_task --example-srt copy audio.mp3 from aeneas/tools/res, rename it as sonnet.mp3 and open it in any media player that supports subtitles (mpv, for example) |
The resulting file (sonnet.srt) will be completely out of sync with sonnet.mp3, but with aeneas v1.7.0 (|pip install aeneas==1.7.0|) it will produce the right timings and correct result.
I don't know if it is a bug or not but aeneas doesn't official supports Windows so I'll just leave it here for someone who tried to install it on Windows x64.
And thank you very much for this groundbreaking software. It's amazing!
kelciour
More precisely, Windows x64 and Python x64. It works fine with Python x86 but audio duration limit is much, much shorter (maybe ~30 minutes) in that case.
Maybe there's something wrong with DTWAligner because I noticed these lines in the log file:
For comprasion, these lines from the exact same file but on Ubuntu where I got correct timings.
How to produce:
The resulting file (sonnet.srt) will be completely out of sync with sonnet.mp3, but with aeneas v1.7.0 (
pip install aeneas==1.7.0) it will produce the right timings and correct result.I don't know if it is a bug or not but aeneas doesn't official supports Windows so I'll just leave it here for someone who tried to install it on Windows x64.
And thank you very much for this groundbreaking software. It's amazing!