/Impulcifer

Measurement and processing of binaural impulse responses for personalized surround virtualization on headphones.

Primary LanguageJupyter NotebookMIT LicenseMIT

Impulcifer

Impulcifer is a tool for creating binaural room impulse responses (BRIR) for speaker virtualization on headphones.

Normally headphones sound inside your head which is a clear disadvantage for games and movies but also for music because basically all material has been created for speakers. Virtual surround technologies for headphones have existed for a some time by now but almost all of them fail to fulfill expectations of out of head sound localization and the naturalness of speakers. This is because your brains have learned to localize sounds only with your ears and head and not with anyone else's. Surround sound on headphones can only be convincing when the surround virtualization technology has been tailored for your ears. BRIR is the tailored model for supreme virtual speaker surround on headphones. When done right virtualized speakers can be indistinguishable from real speakers.

Watch these videos to get an idea what good BRIRs can do. The method used by Smyth Realizer and Creative Super X-Fi demos is the same what Impulcifer uses.

These demos are trying to make headphones sound as much as possible like the speakers they have in the demo room for a good wow effect. Impulcifer actually takes this even further because Impulcifer can do measurements with only one speaker so you don't need access to surround speaker setup and can do room acoustic corrections which are normally not possible in real rooms with DSP.

Installing

Impulcifer is used from a command line and requires some prerequisites. These installation instructions will guide you through installing everything that is needed to run Impulcifer on you own PC.

  • Download and install Git: https://git-scm.com/downloads. When installing Git on Windows, use Windows SSL verification instead of Open SSL or you might run into problems when installing project dependencies.
  • Download and install 64-bit Python 3.8. Make sure to check Add Python 3.8 to PATH.
  • You may need to install libsndfile if you're having problems with soundfile when installing requirements.txt.
  • On Linux you may need to install Python dev packages
sudo apt install python3-dev python3-pip python3-venv

Rest will be done in terminal / command prompt. On Windows you'll find it by searching cmd in Start menu. You should be able to simply copy and paste in these commands.

  • Git clone Impulcifer. This will create a folder in your home folder called Impulcifer. See Updating for other versions than the latest.
git clone https://github.com/jaakkopasanen/Impulcifer
  • Go to Impulcifer folder.
cd Impulcifer
  • Create virtual environment for the project.
python -m venv venv
  • Activate virtualenv.
# On Windows
venv\Scripts\activate
# On Mac and Linux
. venv/bin/activate
  • Update pip and setuptools
python -m pip install -U pip
  • Install required packages.
pip install -U -r requirements.txt
  • Verify installation. You should see help printed if everything went well.
python impulcifer.py --help

When coming back at a later time you'll only need to activate virtual environment again before using Impulcifer.

cd Impulcifer
# On Windows
venv\Scripts\activate
# On Mac and Linux
. venv/bin/activate

Updating

Impulcifer is under active development and updates quite frequently. Take a look at the Changelog to see what has changed.

Versions in Changelog have Git tags with which you can switch to another version than the latest one:

# Check available versions
git tag
# Update to a specific version
git checkout 1.0.0

You can update your own copy to the latest versions by running:

git checkout master
git pull

required packages change quite rarely but sometimes they do and then it's necessary to upgrade them

python -m pip install -U -r requirements.txt

You can always invoke the update for required packages, it does no harm when nothing has changed.

Demo

The actual BRIR measurements require a little investment in measurement gear and the chances are that you're here before you have acquired them. There is a demo available for testing out Impulcifer without having to do the actual measurements. data/demo folder contains five measurement files which are needed for running Impulcifer. headphones.wav has the sine sweep recordings done with headphones and all the rest files are recordings done with stereo speakers in multiple stages.

You can try out what Impulcifer does by running:

python impulcifer.py --test_signal=data/sweep-6.15s-48000Hz-32bit-2.93Hz-24000Hz.pkl --dir_path=data/demo 

Impulcifer will now process the measurements and produce hrir.wav and hesuvi.wav which can be used with headphone speaker virtualization software such as HeSuVi to make headphones sound like speakers in a room. When testing with HeSuVi copy hesuvi.wav into C:\Program Files\Equalizer APO\config\Hesuvi\hrir, (re)start HeSuVi and select hesuvi.wav from the Common HRIRs list on Virtualization tab.

Measurement

BRIR measurements are done with binaural microphones which are also called ear canal blocking microphones or in-ear microphones. Exponential sine sweep test signal is played on speakers and the sound is recorded with the microphones at ear canal openings. This setup ensures that the sound measured by the microphones is affected by the room, your body, head and ears just like it is when listening to music playing on speakers. Impulcifer will then transform these recordings into impulse responses, one per each speaker-ear pair.

Guide for doing the measurements yourself and comments about the gear needed to do it can be found in measurements page of Impulcifer wiki. The whole process is really quite simple and doesn't take more than couple of minutes. Reading through the measurement guide is most strongly recommended when doing measurements the first time or using a different speaker configuration the first time.

Following is a quick reference for running the measurements once you're familiar with the process. If you always use my_hrir as the temporary folder and rename it after the processing has been done, you don't have to change the following commands at all and you can simply copy-paste them for super quick process.

7.1 Speaker Setup

Steps and commands for doing measurements with 7.1 surround system:

Setup Command
Put microphones in ears, put headphones on python recorder.py --play="data/sweep-seg-FL,FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/headphones.wav"
Take heaphones off, look forward python recorder.py --play="data/sweep-seg-FL,FC,FR,SR,BR,BL,SL-7.1-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/FL,FC,FR,SR,BR,BL,SL.wav"

Stereo Speaker Setup

Steps and commands for doing measurements with two speakers in four stages:

Setup Command
Put microphones in ears, put on headphones python recorder.py --play="data/sweep-seg-FL,FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/headphones.wav"
Take heaphones off, look forward python recorder.py --play="data/sweep-seg-FL,FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/FL,FR.wav"
Look 120 degrees left (left speaker should be on your right) python recorder.py --play="data/sweep-seg-FL,FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/SR,BR.wav"
Look 120 degrees right (right speaker should be on your left) python recorder.py --play="data/sweep-seg-FL,FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/BL,SL.wav"
Look directly at the left speaker OR... python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/FC.wav"
...Look directly at the right speaker python recorder.py --play="data/sweep-seg-FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/FC.wav"

Single Speaker

Steps and command for doing measurements with just a single speaker in 7 steps. All speaker measurements use either sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav or sweep-seg-FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav depending if the speaker is connected to left or right cable terminals in the amplifier. These commands assume the speaker is connected to left speaker terminals.

Setup Command
Put microphones in ears, put on headphones python recorder.py --play="data/sweep-seg-FL,FR-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/headphones.wav"
Look 30 degrees right of the speaker (speaker on your front left) python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/FL.wav"
Look directly at the speaker python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/FC.wav"
Look 30 degrees left of the speaker (speaker on you front right) python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/FR.wav"
Look 90 degrees left of the speaker (speaker on your right) python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/SR.wav"
Look 150 degrees left of the speaker (speaker on your back right) python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/BR.wav"
Look 150 degrees right of the speaker (speaker on you back left) python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/BL.wav"
Look 90 degrees right of the speaker (speaker on your left) python recorder.py --play="data/sweep-seg-FL-stereo-6.15s-48000Hz-32bit-2.93Hz-24000Hz.wav" --record="data/my_hrir/SL.wav"

Processing

Once you have obtained the sine sweep recordings, you can turn them into a BRIR file with Impulcifer. All the processing is done by running a single command on command line. The command below assumes you have made a speaker recordings and a headphones recording and saved the recording files into data/my_hrir folder. Start command prompt, jump to Impulcifer folder and activate the virtual environment as described in the installation instructions if you don't have command prompt open yet. Sine sweep recordings are processed by running impulcifer.py with Python as shown below.

python impulcifer.py --test_signal="data/sweep-6.15s-48000Hz-32bit-2.93Hz-24000Hz.pkl" --dir_path="data/my_hrir" --plot

You should have several WAV files and graphs in the folder. hesuvi.wav can now be used with HeSuVi to make your headphones sound like speakers.

--dir_path=data/my_hrir tells Impulcifer that the recordings can be found in a folder called my_hrir under data. Impulcifer will also write all the output files into this folder.

Impulcifer always needs to know which sine sweep signal was used during recording process. Test signal can be either a WAV (.wav) file or a Pickle (.pkl) file. Test signal is read from a file called test.pkl or test.wav. impulse_response_estimator.py produces both but using a Pickle file is a bit faster. Pickle file however cannot be used with recorder.py. An alternative way of passing the test signal is with a command line argument --test_signal which takes is a path to the file eg. --test_signal="data/sweep-6.15s-48000Hz-32bit-2.93Hz-24000Hz.pkl".

Sine sweep recordings are read from WAV files which have speaker names separated with commas and .wav extension eg. FL,FR.wav. The individual speakers in the given file must be recorded in the order of the speaker names in the file name. There can be multiple files if the recording was done with multiple steps as is the case when recording 7.1 setup with two speakers. In that case there should be FL,FR.wav, SR,BR.wav, BL,SL.wav and FC.wav files in the folder.

Room Correction

Similar pattern is used for room acoustics measurements. The idea is to measure room response with a calibrated measurement microphone in the exact same spot where the binaural microphones were. Room measurement files have file name format of room-<SPEAKERS>-<left|right>.wav, where <SPEAKERS> is the comma separated list of speaker names and <left|right> is either "left" or "right". This tells if the measurement microphone is measuring at the left or right ear position. File name could be for example room-FL,FR-left.wav. Impulcifer does not support stereo measurement microphones because vast majority of measurement microphones are mono. Room recording files need to be single track only. Some measurement microphones like MiniDSP UMIK-1 are seen as stereo microphones by Windows and will for that reason record a stereo file. recorder.py can force the capture to be one channel by setting --channels=1.

Generic room measurements can be done for speakers with which it's hard to position the measurement microphone correctly. Impulcifer reads these measurements from room.wav file which can contain any number of tracks and any number of sweeps per track. All the sweeps are being read and their frequency responses are combined. The combined frequency response is used for room correction with the speakers that don't have specific measurements (room-FL,FR-left.wav etc...).

There are two methods for combining the frequency responses: "average" and "conservative". Average method takes the average frequency response of all the measurements and builds the room correction equalization with that. Conservative method takes the absolute minimum error for each frequency and only if all the measurements are on the same side of 0 level at the given frequency. This ensures that there will never be room correction adjustments that would make the frequency response of any of the measurements worse. These methods are available with --fr_combination_method=average and --fr_combination_method=conservative.

Upper frequency limit for room measurements can be adjusted with parameters --specific_limit and --generic_limit. These will limit the room correction equalization to 0 dB above that frequency. This can be useful for avoiding pitfalls of room correction in high frequencies. --specific_limit applies to room measurements which specify the ear and --generic_limit to room measurements which don't. Typically room dominates the frequency response below 300 or 400 Hz and speakers dominate above 700 Hz. Speaker's problems cannot be fixed based on in-room measurements and therefore the limits should usually be placed at 700 Hz. The octave leading up to the limit (eg. 350 to 700 Hz) will be sloped down from full EQ effect (at 350 Hz) to 0 dB at the limit (700 Hz). Other, higher, values can be tried out and they can improve the sound but there are not guarantees about that.

Generic room measurements are not expected to be in the same location as the binaural microphones were so limiting the equalization to less than 1 kHz is probably a good idea. Conservative combination method with several measurements is safer method and with that it should be safer to try to increase the limit up from 1 kHz. For example --specific_limit=5000 --generic_limit=2000 would ensure that room correction won't adjust frequency response of BRIR above 5 kHz for any speaker-ear pairs and above 2 kHz for speaker-ear pairs that don't have specific room measurements.

Room measurements can be calibrated with the measurement microphone calibration file called room-mic-calibration.txt or room-mic-calibration.csv. This must be a CSV file where the first column contains frequencies and the second one amplitude data. Data is expected to be calibration data and not equalization data. This means that the calibration data is subtracted from the room frequency responses. An alternative way of passing in the measurement microphone calibration file is with a command line argument --room_mic_calibration and it takes a path to the file eg. --room_mic_calibration="data/umik-1_90deg.txt"

Room frequency response target is read from a file called room-target.txt or room-target.csv. Head related impulse responses will be equalized with the difference between room response measurements and room response target. An alternative way to pass in the target file is with a command line argument --room_target eg. --room_target="data/harman-in-room-loudspeaker-target.csv".

Room correction can be skipped by adding a command line argument --no_room_correction without any value.

Headphone Compensation

Impulcifer will compensate for the headphone frequency response using headphone sine sweep recording if the folder contains file called headphones.wav. If you have the file but would like not to have headphone compensation, you can simply rename the file for example as headphones.wav.bak and run the command again.

Headphone equalization can be baked into the produced BRIR file by having a file called eq.csv in the folder. The eq file must be an AutoEQ produced result CSV file. Separate equalizations for left and right channels are supported with files eq-left.csv and eq-right.csv. Headphone equalization is useful for in-ear monitors because it's not possible to do headphone compensation with IEMs. When using IEMS you still need an around ear headphone for the headphone compensation. eq.wav is no longer supported!

Impulcifer will bake the frequency response transformation from the CSV file into the BRIR and you can enjoy speaker sound with your IEMs. You can generate this filter with AutoEQ; see usage instructions for using sound signatures to learn how to transfer one headphone into another. In this case the input directory needs to point to the IEM, compensation curve is the curve of the measurement system used to measure the IEM and the sound signature needs point to the existing result of the headphone which was used to make the headphone compensation recording.

For example if the headphone compensation recording was made with Sennheiser HD 650 and you want to enjoy Impulcifer produced BRIR with Campfire Andromeda, you should run:

python frequency_response.py --input_dir="oratory1990/data/inear/Campfire Audio Andromeda" --output_dir="my_results/Andromeda (HD 650)" --compensation="compensation/harman_over-ear_2018_wo_bass.csv" --sound_signature="results/oratory1990/harman_over-ear_2018/Sennheiser HD 650/Sennheiser HD 650.csv" --equalize --bass_boost=4 --max_gain=6

and then copy AutoEq/my_results/Andromeda (HD 650)/Campfire Audio Andromeda.csv to Impulcifer/data/my_hrir/eq.csv.

See how the Harman over-ear target is used for IEM in this case. This is because the goal is to make Andromeda sound as similar as possible to HD 650, which is an over-ear headphone. Normally with AutoEQ you would use Harman in-ear target for IEMs but not in this case.

Headphone compensation can be skipped by adding a command line argument --no_headphone_compensation without any value.

Sampling Rate

Outputs with different sampling rates than the recording sampling rate can be produced with --fs parameter. This parameter takes a sampling rate in Hertz as value and will then resample the output BRIR to the desired sampling rate if the recording and output sampling rates differ. For example --fs=44100.

Plotting Graphs

Various graphs can be produced by providing --plot parameter to Impulcifer. These can be helpful in figuring out what went wrong if the produced BRIR doesn't sound right. Producing the plots will take some time.

  • pre plots are the unprocessed BRIR measurement
  • room plots are room measurements done with measurement microphone
  • post plots are the final results after all processing

Channel Balance Correction

Channel balance can be corrected with --channel_balance parameter. In ideal case this would not be needed and the natural channel balance after headphone equalization and room correction would be perfect but this is not always the case since there are multiple factors which affect that like placement of the binaural microphones. There are six different strategies available for channel balance correction.

Setting --channel_balance=trend will equalize right side by the difference trend of left and right sides. This is a very smooth difference curve over the entire spectrum. Trend will not affect small deviations and therefore doesn't warp the frequency response which could lead to uncanny sensations. Bass, mids and treble are all centered when using trend. Trend is probably the best choice in most situations.

Setting --channel_balance=mids will find a gain level for right side which makes the mid frequencies (100, 3000) average level match that of the left side. This is essentially an automatic guess for the numeric strategy value.

Setting --channel_balance=1.4 or any numerical value will amplify right side IRs by that number of decibels. Positive values will boost right side and negative values will attenuate right side. You can find the correct value by trial and error either with Impulcifer or your virtualization software and running Impulcifer again with the best value. Typically vocals or speech is good reference for finding the right side gain value because these are most often mixed in the center and therefore are the most important aspect of channel balance.

Setting --channel_balance=avg will equalize both left and right sides to the their average frequency response and --channel_balance=min will equalize them to the minimum of the left and right side frequency response curves. Using minimum instead of average will be better for avoiding narrow spikes in the equalization curve but which is better in the end varies case by case. These strategies might cause uncanny sensation because of frequency response warping.

--channel_balance=left will equalize right side IRs to have the same frequency response as left side IRs and --channel_balance=right will do the same in reverse. These strategies might cause uncanny sensation because of frequency response warping.

Level Adjustment

Output BRIR level can be adjusted with --target_level parameter which will normalize the BRIR gain to the given numeric value. The level is calculated from all frequencies excluding lowest bass frequencies and highest treble frequencies and then the level is adjusted to the target level. Setting --target_level=0 will ensure that BRIR average gain is about 0 dB. Keep in mind that there often is large variance in the gain of different frequencies so target level of 0 dB will not mean that the BRIR would not produce clipping. Typically the desired level is several dB negative such as --target_level=-12.5. Target level is a tool for having same level for different BRIRs for easier comparison.

Decay Time Management

The room decay time (reverb time) captured in the binaural room impulse responses can be shortened with --decay parameter. The value is a time it should take for the sound to decay by 60 dB in milliseconds. When the natural decay time is longer than the given target, the impulse response tails will be shortened with a slope to achieve the desired decay velocity. Decay times are not increased if the target is longer than the natural one. The decay time management can be a powerful tool for controlling ringing in the room without having to do any physical room treatments.

Contact

Issues are the way to go if you are experiencing problems, have ideas or if there is something unclear about how things are done or documented.

You can find me in Reddit and Head-fi if you just want to say hello.

There is also a Head-fi thread about Impulcifer.