auphonic.com - Latest entries

Podcast Comparison, Part 4: MPEG-4 iTunes-style Metadata (AAC Audio, M4A, MP4)

grh@auphonic.com (grh) — Sat, 18 Feb 2012 09:35:40 +0100

The last metadata format in this comparison series will be metadata for AAC audio files, which are usually included in an MP4 container (see MPEG-4 Part 14). MP4 audio files can have various file extensions, most commonly m4a, but also mp4 or m4b (for audio books).
So let the confusion continue ...

MP4, AAC and iTunes-style Metadata

The MPEG-4 container format is based on Apple's QuickTime container and each MPEG-4 file must have a major file brand. For example, an AAC audio file typically lists M4A as its major file brand. See AtomicParsley for a description of the MPEG-4 container and MP4 Registration Authority for a list of official file brand identifiers.

Unfortunately no official metadata or tagging standard exists for MP4. The MPEG-4 specification just defines a standard place to put metadata, but there is no standard on any fields, except cprt for copyright :).
Therefore different metadata styles are possible: iTunes-style metadata, 3gp-style metadata, MP4 ID3v2 tags and more. Most podcasts and audio files use iTunes-style metadata.

So far so good, but iTunes-style metadata isn't defined in any publicly available document. Its format is determined by the types of files that iTunes and the iTunes Music Store produce and provide (see AtomicParsley, MPEG4 Files, bottom).
Some open source projects assembled a list of currently known iTunes tags: iTunes Metadata (mp4v2) and Known iTunes Metadata Atoms (AtomicParsley)

Helpful tools to analyze MP4 audio are mp4v2 (including the tools mp4chaps, mp4art, mp4file, mp4info, mp4track) and faad, the Free Advanced Audio Decoder, can list AAC codec details and metadata tags.

Used Tags

But now let's look at some real life MP4 files. The following table lists iTunes-style metadata tags and the encoding/tagging software of a few podcasts:

Podcast	Encoder	Used Tags
Elementarfragen 9	GarageBand	name, artist, composer, encoded with, date, album, genre (Podcast), comments (for URL), description, lyrics (for show notes)
Fanboys Podcast 49	Hindenburg Journalist	name, artist, encoded with, album, track, genre (Podcast), comments (for show notes), description (for show notes), lyrics (for show notes)
Engadget Podcast 270	GarageBand	name, artist, encoded with, date, album, track, genre (Podcast), comments (for URL)
Newz of the world 1	iTunes	name, artist, encoded with, date, album, track, genre (Podcast), BPM, album artist
Naked Scientists	unknown	name, artist, encoded with, date, album, description (for show notes), genre (Podcast), BPM
Pressrecord 16	GarageBand	name, artist, composer, encoded with, album, comments (for show notes)
Disney Land News Today	GarageBand	name, encoded with
Inside Home Recording	GarageBand	name, artist, composer, encoded with, date, album, track, genre (Podcast), lyrics (for show notes)
Fubar Podcast 6	Hindenburg Journalist	name, artist, encoded with, album, genre (Podcast), album artist, lyrics (for show notes)

AAC Details, Chapter Marks and Images

It's also interesting to analyze which AAC flavours (LC or HE AAC, for a short description see AAC Details) and bitrates are used.
Furthermore we compared mostly Enhanced Podcasts, which contain chapter marks. And there are of course multiple ways to embed chapter marks in an MP4 audio stream (e.g. QuickTime chapters, Nero chapters), but basically all podcasts use QuickTime chapters.

The next table shows AAC details, bitrates, cover image size, number of chapter marks and length of the analyzed podcasts:

Podcast	Format	Image	Chapter Marks	Length
Elementarfragen 9	LC AAC, 128kbps	jpeg, 1024x1024	15	02:04:22
Fanboys Podcast 49	HE AAC, 56kbps	png 320x320, jpeg 1024x1024	24	01:08:17
Engadget Podcast 270	LC AAC, 80kbps	jpeg, 300x300	6	01:26:35
Newz of the world 1	HE AAC, 49kbps	jpeg, 600x600	6	57:52
Naked Scientists	LC AAC, 48kbps	jpeg, 300x300	26	01:07:32
Pressrecord 16	LC AAC, 128kbps	jpeg, 300x300	No	01:20:48
Disney Land News Today	LC AAC, 192kbps	jpeg, 300x300	4	01:58:22
Inside Home Recording	LC AAC, 128kbps	jpeg, 300x300	8	01:02:08
Fubar Podcast 6	HE AAC, 80kbps	jpeg, 1200x1200	33	04:04:30

Podcast Comparison Series

If you are interested in more properties of popular podcasts, look through the other posts of the podcast comparison series:

Podcast Comparison, Part 3: Ogg Vorbis Metadata (Vorbis Comment)

grh@auphonic.com (grh) — Sun, 22 Jan 2012 19:50:59 +0100

A new week - a new format: this weeks post is about metadata in Ogg Vorbis audio. The metadata container for these files is called Vorbis comment, which is also used in the FLAC, Theora and Speex file formats.
For a general introduction into audio metadata see e.g. the previous post in this comparison series: Part 2: MP3 Metadata (ID3 Tags).

The tool to analyze and create metadata for Ogg Vorbis files is called vorbiscomment.

About Vorbis Comment

Vorbis comment is a relatively simple metadata format specified by Xiph.org, similar to ID3 tags for MP3 files. It allows the user to set arbitrary fieldname/value pairs, using UTF-8 Unicode as the character encoding. Field names can be used several times and are case insensitive.

The exact format specification and a minimal list of standard fields is located at Vorbis comment format specification. Further field recommendations and details can be found at the Xiph.org Vorbis comment Wiki and at Ogg Vorbis Comment Field Recommendations.

Used Field Names

In the following table the bitrate, channels and used field names of various Ogg Vorbis podcasts are listed:

Podcast	Bitrate	Channels	Used Field Names
Alternativlos 21	74.9kbps	Stereo	ALBUM, ARTIST, COMMENT, GENRE (Speech), TITLE, DATE, TRACKNUMBER, METADATA_BLOCK_PICTURE
Ubuntu UK Podcast (high)	59.5kbps	Stereo	title, artist, genre (Podcast), date, album, TRACKNUMBER, COMMENT, COVERARTMIME, COVERART
Ubuntu UK Podcast (low)	36.2kbps	Mono	title, artist, genre (Podcast), date, album, TRACKNUMBER, COMMENT, COVERARTMIME, COVERART
Radio Tux	96.5kbps	Stereo	title, artist, album, date, copyright, license (for URL), coverartmime, coverartdescription, coverart
Radio New Zealand Afternoons	49.7kbps	Mono	DESCRIPTION, title, artist, album
Pentacast 42	65.1kbps	Stereo	ALBUM, ARTIST, DATE, ENCODER, ENCODING, GENRE, TITLE, TRACKNUMBER
This Week in Tech 205	65.1kbps	Mono	ALBUM, ARTIST, COMMENT (for URL), GENRE (Podcast), TITLE
TuxRadar Season 3, 24	79.5kbps	Stereo	ALBUM, TITLE, TRACKNUMBER, COMMENTS, Authors, ARTIST, DATE, GENRE (Podcast)
Netzpolitik Podcast 114	119kbps	Mono	ALBUM, TITLE, DATE, ARTIST, COMMENTS
An American Atheist 63	98.2kbps	Stereo	ALBUM, ARTIST, DATE, GENRE (News), TITLE, TRACKNUMBER

It's interesting that some podcasts use the field name COMMENTS instead of COMMENT, although the official recommendation is COMMENT and all tools I tested just support COMMENT.
The very popular This Week in Tech (TWiT) podcast offered an Ogg Vorbis version until August 2009. They stopped it because it took too much time and energy to encode all the versions of TWiT (according to TWiT history).

Cover Art

There are two ways to embed jpeg or png pictures within Vorbis comments: the recommended way uses the tag METADATA_BLOCK_PICTURE and the unofficial method the tag COVERART, for details see Vorbis comment, Cover art.
However, some of the tested audio players don't support the recommended version (for example VLC 1.1.12), wherease some support both ways.

The following table shows details about embedded images:

Podcast	Image Tags	Type	Size	Resolution
Alternativlos 21	METADATA_BLOCK_PICTURE	image/png	31.6 KB	521x491
Ubuntu UK Podcast	COVERARTMIME, COVERART	image/jpeg	12.3 KB	388x388
Radio Tux	coverartmime, coverartdescription, coverart	image/jpeg	3.1KB	160x160

The most important observation is, that very few podcasts use cover images at all. Maybe this is a consequence of the two different ways to embed pictures.
Furthermore images should not be too big, otherwise old hardware players might have problems if the Vorbis comment header gets too large.

Podcast Comparison Series

If you are interested in more properties of popular podcasts, look through the other posts of the podcast comparison series:

Podcast Comparison, Part 2: MP3 Metadata (ID3 Tags)

grh@auphonic.com (grh) — Tue, 17 Jan 2012 08:38:15 +0100

The second post of this comparison series is about metadata in MP3 files. Metadata allows information such as the title, artist, comments, cover image and other information about the audio to be stored in the file itself. MP3 files use ID3 tags and in the following I will compare common tags, image details and ID3 versions of various popular podcasts.
The analyzed files are the same as in Podcast Comparison, Part 1: File Formats and Bitrates.

Nice open source tools to analyze MP3 metadata are for instance eyeD3, soxi, mp3diags or EasyTag.

ID3 Versions and Used Tags

There are two major versions of ID3 tags: ID3v1 and ID3v2, with various minor versions. The most popular version implemented today is ID3 version 2.3 (according to the official ID3 site), the latest one is ID3v2.4. For a list and explanation of possible ID3v2 tags see Declared ID3v2 frames.

Here is the list of podcasts with ID3 version, used tags and its encoder software:

Podcast	ID3	Used Tags	Encoder
CRE 188	v2.2	title, artist, album, year, genre (Podcast), Comment (for URL), TPE2 (accompaniment), TENC (encoder software)	iTunes
techzing 162	v2.4	title, artist, album, composer, year, genre (Podcast), track, Comment (for URL)	unknown
Hoaxilla 70	v2.4	title, artist, album, year, genre (Podcast), TPUB (publisher), Comment (for show notes), TSSE (software for encoding)	Lavf52
Planet Money Podcast	v1.0	title, artist, year, genre (Blues)	unknown
Startups For the Rest of Us 61	v2.2	title, artist, album, year, genre (Podcast), Comment (description), Lyrics (description), TENC (encoder software)	iTunes
Alternativlos 21	v2.3	title, artist, album, year, track, genre (Speech), Comment (description)	unknown
Bits und so 272	v2.2	title, artist, album, year, track, genre (Podcast), Comment (for URL), TENC (encoder software)	iTunes
Elementarfragen 9	v2.2	title, artist, album, year, genre (Podcast), Comment (for URL), Lyrics (for show notes), TSSE (encoding software), TDAT (date), TBPM (BPM), TIT3 (subtitles)	GarageBand
Ubuntu UK Podcast	v2.4	title, artist, album, year, track, genre (Podcast), Comment (for show notes), Lyrics (for show notes), TDTG (tagging time)	unknown
Fanboys Podcast 49	v2.3	title, artist, album, year, TENC (encoder software)	Hindenburg Journalist
This American Life 420	v2.2	title, artist, year, genre (Podcast), Comment (for copyright information)	unknown
ARD Radio Tatort	v2.3	title, artist, album, year, genre (Podcast), TPUB (publisher), Comment (for show notes), link frames (WCOP, WOAF, WOAS, WORS, WPUB, WXXX), TEXT (Text writer), TLAN (language), TRSN (radio station), TSRO (radio station owner)	unknown
Deutschland Radio Wissen	v2.3	title, artist, album, year, TCOP (copyright), link frames (WCOP, WOAF, WORS, WPUB)	unknown
Freakonomics	v2.3	title, artist, album, year, genre (Audio), Comment, WXXX (for URL), TCOP (copyright), TENC (encoder software), TXXX (tagging time)	DigaSystem DigaPorter
OE1 Geschichte Podcast	v2.3	title, artist, album, year, genre (Podcast), Comment, TCOP (copyright), TENC (encoder software), link frames (WCOP, WORS, WPUB), TOFN (original filename)	lame
Wait Wait ... don't tell me	v1.0	title, artist, year, genre (Blues)	unknown
Radio New Zealand Afternoons	v2.3	title, artist, album, Comment (for show notes), TSSE (software settings for encoding), TLEN (length)	lame
Inside Home Recording 97	v2.2	title, artist, album, year, track, genre (Podcast), BPM, Lyrics (for show notes), TSSE (software settings for encoding)	GarageBand
Radio Tux	v2.4	title, artist, album, year, genre (Podcast), WXXX (for URL)	unknown

It can be seen that iTunes and GarageBand use ID3v2.2 tags, although the recommended version is ID3v2.3 (as implemented by most public radio stations).
Mp3diags even displays a warning for v2.2 and v2.4 tags:
"No ID3v2.3.0 tag found, although this is the most popular tag for storing song information."

Nearly all podcasts use the genre "Podcast", which is not even included in the official ID3 genre list. Furthermore most unknwon encoders are probably lame.

Attached Picture

It is possible to attach pictures to an MP3 file. These images should be in jpeg or png format, they have a defined picture type (other, front cover, artist, conductor, composer, etc.) and might contain an additional description - see ID3 Attached Picture.

Here are the details about attached pictures of all analyzed podcasts which include an image:

Podcast	Picture Type	MIME Type	Size	Resolution
CRE 188	OTHER Image	image/JPG	238 KB	600x600
techzing 162	OTHER Image	image/png	473.8 KB	600x600
Hoaxilla 70	OTHER Image	image/jpeg	98.4 KB	600x600
Startups For the Rest of Us 61	OTHER Image	image/PNG	24.9 KB	271x127
Alternativlos 21	FRONT_COVER Image	image/png	31.7 KB	521x491
Elementarfragen 9	OTHER Image	image/PNG	55.7 KB	1024x1024
Ubuntu UK Podcast	FRONT_COVER Image	image/jpeg	12.3 KB	388x388
Fanboys Podcast 49	OTHER Image	image/jpeg	13.4 KB	300x300
This American Life 420	OTHER Image	image/JPG	20 KB	300x300
ARD Radio Tatort	OTHER Image	image/jpg	33.8 KB	300x300
Freakonomics	FRONT_COVER Image	image/jpeg	45 KB	300x300
OE1 Geschichte Podcast	FRONT_COVER Image	image/png	153.4 KB	300x300
Inside Home Recording 97	OTHER Image	image/JPG	68 KB	300x299
Radio Tux	OTHER Image	image/jpeg	3.2 KB	160x160

One small detail: iTunes and GarageBand use capital letters for mimetypes (e.g. image/PNG instead of image/png). This confuses some tag readers and players, for instance the EasyTag editor is not able to display these images.

Podcast Comparison Series

If you are interested in more properties of popular podcasts, look through the other posts of the podcast comparison series:

Podcast Comparison, Part 1: File Formats and Bitrates

grh@auphonic.com (grh) — Sat, 07 Jan 2012 16:55:46 +0100

This is the first entry out of a series of posts, where I compare various aspects of popular podcasts. Part 1 is about file formats, bitrates, samplerates, channels, size, length and filenames.

The selection is based on Flattr podcast charts (mostly german), iTunes Store Top 10 Podcasts (english, german), podcasts by public radios (english, german) and some random ones to get more diversity.

List of Analyzed Podcasts

Here is the list of podcasts with links to the analyzed episodes, the available file formats and filenames (only for one format):

Podcast	Filename of Episode	File Formats
CRE 188	cre188-telecomix.mp3	MP3
techzing 162	techzing-162.mp3	MP3
Hoaxilla 70	hoaxcast070_wahrsager.mp3	MP3
Planet Money Podcast	20111230_blog_pmoney.mp3	MP3
Startups For the Rest of Us 61	startups-for-the-rest-of-us-061.mp3	MP3
Alternativlos 21	alternativlos-21.mp3	MP3, OGG
Bits und so 272	bits-2011-12-23.mp3	MP3, M4A
Elementarfragen 9	ef09mp.mp3	MP3, M4A
Ubuntu UK Podcast	uupc_s04e22_high.mp3	2xMP3, 2xOGG
Fanboys Podcast 49	FAN049.m4a	MP3, M4A
This American Life 420	420.mp3	MP3
ARD Radio Tatort	2011_podcast_unter_verdacht.mp3	MP3
Deutschland Radio Wissen	drw_201112301745_die_wissensnachrichten_ des_tages_79e879d8.mp3	MP3
Freakonomics	freakonomics_podcast122111.mp3	MP3
OE1 Geschichte Podcast	OE1_geschichte_111230.MP3	MP3
Wait Wait ... don't tell me	20111231_waitwait_04.mp3	MP3
Radio New Zealand Afternoons	aft-20111223-1606-the_panel_christmas_special-00.ogg	MP3, OGG
Inside Home Recording 97	IHR_2011_12_05.m4a	MP3, M4A
Radio Tux	2011-12-23.RadioTux.WeihnachtsSpecial.2011.ogg	MP3, OGG

It's easy to see that the by far most common file format is MP3, for an overview of file formats see Audio File Formats for Podcasts. In addition to the MP3 version, some podcasts offer Ogg Vorbis (file extension OGG, strong in open source community) and AAC (file extension M4A, strong in apple community).

Some filenames are also very unintuitive and long. That's especially annoying on MP3 players with small displays, because it can be very hard to find a specific podcast episode.

Audio File Format Information

The next table lists more detailed information about the used audio codec, samplerate, bitrate, channels, size and length:

Podcast	Audio Codec	Samplerate	Bitrate	Channels	Size	Length
CRE 188	MP3	44100	96k CBR	Joint Stereo	72.4M	1:40:15
techzing 162	MP3	44100	128k CBR	Joint Stereo	108M	1:51:48
Hoaxilla 70	MP3	22050	64k CBR	Joint Stereo	25.4M	52:36
Planet Money Podcast	MP3	44100	64k CBR	Mono	9.98M	20:47
Startups For the Rest of Us 61	MP3	44100	160k CBR	Joint Stereo	43.58M	36:14
Alternativlos 21	MP3	44100	115k VBR	Joint Stereo	82.2M	1:35:39
Alternativlos 21	Ogg Vorbis	44100	74.9k VBR	Joint Stereo	53.7M	1:35:39
Bits und so 272	MP3	44100	64k CBR	Mono	63.2M	2:11:40
Elementarfragen 9	MP3	44100	128k CBR	Simple Stereo	119M	2:04:23
Elementarfragen 9	LC AAC	44100	128k CBR	Joint Stereo	121.7M	2:04:23
Ubuntu UK Podcast	MP3 high	44100	126k VBR	Simple Stereo	63.5M	1:07:17
Ubuntu UK Podcast	MP3 low	44100	32k CBR	Mono	16.1M	1:07:17
Ubuntu UK Podcast	Ogg Vorbis high	44100	59.5k VBR	Joint Stereo	30.1M	1:07:17
Ubuntu UK Podcast	Ogg Vorbis low	44100	36.2k VBR	Mono	18.2M	1:07:17
Fanboys Podcast 49	MP3	44100	64k CBR	Mono	32.8M	1:08:17
Fanboys Podcast 49	HE AAC	44100	56k CBR	Joint Stereo	29M	1:08:17
This American Life 420	MP3	44100	64k CBR	Mono	28.2M	58:40
ARD Radio Tatort	MP3	44100	128k CBR	Simple Stereo	51.8M	53:57
Deutschland Radio Wissen	MP3	44100	64k CBR	Mono	2.92M	06:06
Freakonomics	MP3	44100	96k CBR	Simple Stereo	22.1M	30:36
OE1 Geschichte Podcast	MP3	44100	64k CBR	Mono	2.50M	04:52
Wait Wait ... don't tell me	MP3	44100	64k CBR	Mono	5.33M	11:05
Radio New Zealand Afternoons	MP3	22050	48k CBR	Mono	18.7M	51:53
Radio New Zealand Afternoons	Ogg Vorbis	48000	49.7k VBR	Mono	19.3M	51:53
Inside Home Recording 97	MP3	44100	128k CBR	Joint Stereo	59.7M	01:02:08
Inside Home Recording 97	LC AAC	44100	128k CBR	Joint Stereo	62.4M	01:02:08
Radio Tux	MP3	44100	128k CBR	Joint Stereo	163M	02:49:26
Radio Tux	Ogg Vorbis	44100	96.5k VBR	Joint Stereo	123M	02:49:26

Suprisingly, nearly all MP3 and AAC podcasts use constant bitrate coding (CBR), Ogg Vorbis podcasts variable bitrate coding (VBR). The bitrates range from 32kbps up to 160kbps.
In CBR encoding, the bitrate is kept constant across the entire file, whereas VBR coding tries to maintain the quality during the whole stream by choosing the optimal amout of data to represent each audio frame. VBR will almost always produce equal or better perceived quality results than CBR for files of the same size or average bitrate (from hydrogen), however, some hardware based MP3 decoders don't reliably support VBR.

Furthermore there are two ways to represent stereo audio in lossy codecs: in simple stereo each channel is treated as a completely separate audio signal, whereas joint stereo supports more complex methods of stereo coding for the sake of efficiency.
Joint stereo is very widespread in MP3 podcasts and basically all Ogg Vorbis/AAC files use it.

Podcast Comparison Series

If you are interested in more properties of popular podcasts, look through the other posts of the podcast comparison series:

Podcast Barcamp Vienna: Tools and Enhanced Podcasts

grh@auphonic.com (grh) — Sun, 23 Oct 2011 21:01:21 +0200

Last weekend, 15. + 16. October 2011, the first austrian podcast barcamp (#pccvie) took place in the charming sektor5 coworking space in Vienna. We had really inspiring discussions with many local podcasters and technicians, for some detailed reports in german see Podcastcamp 2001 by Nico Grienauer and Podcast Barcamp: Ein tweetiger und booiger Rückblick by Robert Lender.
I just want to add some thoughts to the discussion about enhanced podcasts and will list some practical tools.

Enhanced Podcasts

It seems that more and more podcasters try to create enhanced podcasts, where additional metadata like chapter marks, hyperlinks or images, are added to the audio file.
However, just a few players display chapter marks correctly and there are no easy-to-use tools to create enhanced podcasts - that's what we want to change :) !

At the moment enhanced podcasts exist mainly in the AAC file format (see also Audio File Formats for Podcasts). It is possible to create chapter marks with the mp4chaps tool, which is included in the open source mp4v2 library, and players on "i"-devices (Quicktime Player, iTunes, iPod, iPhone, etc.) should support them.

For MP3 an ID3v2 Chapter standard exists, but besides the ID3v2 Chapter Tool (developed by BBC open source) nobody seems to use it. With Ogg Vorbis it's even worse, AFAIK there is no similar standard.

We will try to support all existing standards and as many file formats as possible, some more research will be necessary to find a good solution for Ogg Vorbis.
As a fallback, chapter marks could be added to the lyrics tag, so that players without enhanced podcast support show at least something (as suggested by Thomas Perl from gPodder).

Tools for Podcast Producers

Some popular open source tools for podcast producers:

Audacity: A widespread audio editor for recording and editing sounds. Some love it, some not ...
Ardour: The open source DAW (digital audio workstation) for Linux and OS X. Much more advanced than Audacity (e.g. non-destructive editing), but with a steeper learning curve.
Mumble: An alternative to Skype for recording remote podcasters. It uses low-latency, high quality codecs and it's possible to record each voice to a separate channel.
archive.org for hosting: You can use archive.org to host your audio files and save server costs. Nice idea!
Open Cast Matterhorn: A program to manage audio and video content. More targeted to educational institutions, but might be also useful for some podcasters.
mp3diags: Identifies issues in MP3 files, e.g. broken tags, headers, character encoding, and tries to fix them. A nice tool to test your final MP3 files!
Tesseract OCR: An open source optical character recognition (OCR) library, which could be used to extract text out of screencasts or slides. This could serve as additional (searchable!) metadata.

Conclusion

Many thanks to the organizers (Thomas Lohninger, Pepi Zawodsky and Daniel Meßner) for this great event! Barcamp recordings will be online soon - and will serve as first test data for auphonic :)

I hope that the planned Podcast Stammtisch and mailing list will happen in future to keep up the good discussions - otherwise we will hopefully meet again next year at the second podcast barcamp!

The auphonic team

grh@auphonic.com (grh) — Thu, 22 Sep 2011 14:26:18 +0200

After a long, nice summer and few blog posts, it is time to present the current team behind auphonic. At the moment we are two core developers and maintain strong collaborations with three institutions here in Graz.

Enjoy ;)

Core Team

Georg Holzmann

Georg is an audio engineer with passion for machine learning, signal processing, everything related to audio and computer music, open source software, web development and much more.

Discussions and repeating questions from podcasters led him to the idea of an automatic audio post production web service, this was the birth of auphonic.

Previously he worked at scientific institutions (machine learning, audio processing), as a web developer and realized many technology driven, artistic projects.
For more information visit http://grh.mur.at.

André Rattinger

André is a student of Software Development and Business Management at Graz University of Technology. Besides studying he is an enthusiastic python and web developer and a passionate coffee drinker.

If he can't be found in front of his computer, he is probably roasting coffee or trying out a new recipe.

Before auphonic he worked on different python open source projects, including user interface design and web development.

Partners

Science Park Graz

The Science Park Graz is the local version of the Austrian high-tech incubator network AplusB (Academia plus Business).
These incubators are linked to academic institutions and they support the development of technology-based projects for a period of 1.5 years with a free office space, some small financial aids, support in non-technical areas and a close collaboration with scientific partners here in Graz.
Special thanks to Otmar Kühner, Bernhard Weber, Sonja Buchegger and Dagmar Böhm for their ongoing support and help!

Institute of Electronic Music and Acoustics (IEM)

The Institute of Electronic Music and Acoustics (IEM) is a multidisciplinary research center within the Graz University of Music and Performing Arts. Research activities are concentrated mainly in digital signal processing, audio engineering and psycho acoustics, e.g. projects in analysis and syntheses of sound and loudness perception.
The IEM supports our research in signal processing and music information retrieval, many thanks to Dr. Alois Sontacchi for his time and valuable inputs.

Dept. Social Learning at CIS TU Graz

The Abteilung Vernetztes Lernen (Department of Social Learning), at the Graz University of Technology, performs research in the areas of E-Learning, Computer Supported Collaborative Work, Social Learning, Podcasts, Open Content etc.
They are recording lectures at the TU Graz and built knowledge in optimizing post production workflows, which should be further enhanced in the collaboration with auphonic.
These guys are very open minded and supportive - special thanks to Ypatios Grigoriadis and Walther Nagler from the Podcasting team!

Loudness Normalization and Compression of Podcasts and Speech Audio

grh@auphonic.com (grh) — Mon, 25 Jul 2011 23:31:50 +0200

After recording a podcast or speech audio, it is usually necessary to modify the recorded levels. This post illustrates how to normalize the subjective loudness and how to compress the dynamic range (= difference between the loudest and softest sounds) of an audio file.

Loudness normalization is one of the most common misunderstandings in audio post production. Many people use peak normalization, which ensures that the maximum peak (= the maximum value of the audio data) reaches a specific level. However, the human perception of loudness does NOT depend on peak levels, therefore peak normalization is mostly useless. Recordings should be normalized according to its loudness, and not its peak level (see also 10 Myths About Normalization).

The correct calculation of the perceived loudness is actually not that easy, because psychoacoustic properties of human perception must be considered (see e.g. equal-loudness contour, but that's a topic for another blog post). A very rough approximation is the RMS value (= short time quadratic mean of audio data), or even better: use your ears!

There are several reasons for loudness normalization and compression:

Mid-term loudness:
Different speakers or different regions in a recording might have very unequal loudness levels (e.g. a question from the audience at a conference). The audio engineer should balance the loudness of unequal parts, so that a listener doesn't have to adjust the volume all the time.
Short-term loudness:
Speech has a very high dynamic range, it consists of very quiet and very loud parts. However, when listening to podcasts on a portable player, in a car or even in the living room, a more uniform volume is preferred, otherwise consonants like "p", "t" or a burst of laughter might be painful for your ears.
Dynamic range compression (compressors) should be used to make the loud parts more quiet and the quiet parts louder.
Global Loudness:
The overall loudness of your recording should be comparable to previous episodes, to similar podcasts, to radio shows etc.
Peak Levels:
In digital recordings, the maximum peak level must not be louder than the maximum allowed level, otherwise an awful distortion called clipping occurs. Therefore one must ensure that all peaks are under a specific threshold.

In the following I will briefly describe how to adjust your peak levels, the mid-term, short-term and global loudness. The last section presents practical and detailed instructions, how all steps can be reproduced with the open source audio editor Audacity and some free plugins (on Linux, Windows and Mac).

1. Mid-Term: Loudness Normalization

The loudness of different regions in a recording should be unified, so that a listener doesn't have to adjust the volume all the time.
In live recordings this is often done by a sound engineer, using the faders of an audio mixer. She/he just listens to the signal and tries to keep a constant loudness. If the volume of a speaker is too low, the engineer will move the faders up and therefore increases the volume. However, an engineer must react fast enough on volume changes.

In post production, one can use volume envelopes to simulate the same behavior in a digital audio editor. The following picture shows volume envelope curves in Audacity:

Using volume envelopes it is possible to amplify or lower the volume of different regions and speakers, so that in the end all parts have a similar loudness. This has the advantage, that volume fades can be positioned in an optimal way, but it might be a time consuming process if you are not used to it.

Additionally one should take care that background noises or e.g. loud breathing of speakers are not amplified too much.

2. Short-Term: Loudness Range Compression

A compressor reduces the volume of short and loud spikes (consonants like "p", "t", or laughter) and the output audio will have a more uniform loudness. All sounds with levels above the threshold will be reduced. The amount of gain reduction is determined by the ratio: e.g. if the ratio is 2:1 and the input level is 2dB over the threshold, the output level will be just 1dB over the threshold (for a detailed parameter description see e.g. Compressor controls and features).
The following picture illustrates threshold and ratio (picture from MediaCollege):

The threshold should be located above the average loudness of your recording, but below the volume of very loud spikes. For example, if the average level is -20dB, the threshold might be in the range of -16dB to -12dB, depending on the loudness range of the material.

Other compressor parameters can be set to standard values for speech, e.g. ratio = 5:1, attack = 5ms, release = 120ms. Feel free to experiment!

3. Global Level Adjustment

Now one should compare the loudness of the normalized audio file to the loudness of other recordings like similar podcasts or radio shows. Raise or lower your volume, until the levels are similar.

After doing this a few times, you will have a reference audio file (maybe one of your old shows) and can compare new recordings just to that one file.

4. Peak Limiting

The final step is the peak limiter. It ensures that the maximum audio peak level is small enough, so that no clipping or distortion is possible. You should keep your peaks at least below -1 dBFS, some suggest much more (-3dBFS, -6dBFS, ...).
That's because true peaks are often between samples (intersample peaks) and data compression algorithms (MP3, OGG, AAC, ...) produce artifacts at high peak levels.

Tools

All necessary operations should be possible with any modern audio editor.
In the following I will describe how to reproduce all steps with the free and open source audio editor Audacity. Some LADSPA audio plugins are also required, because the current version of Audacity does not include a sufficient compressor and limiter.
On Windows or Mac install the LADSPA plugins bundle as described at Audacity Plugins, on Linux install the swh-plugins (apt-get install swh-plugins on Ubuntu/Debian).

Volume Envelopes:
Select the Envelope Tool from the Tools Toolbar, for more information see Using the Envelope Tool or Audacity Manual Envelope Tool.
Loudness Measurement (optional):
You can use the Contrast Tool to calculate the average level (under Analyze -> Contrast). Select your audio and measure the volume, then you get a value in dB. This might help you while searching for the right compressor threshold.
Compressor:
A good compressor is the SC4 (or SC4 mono) compressor by Steve Harris (LADSPA plugin), you find it somewhere in Effects -> Plugins.
Set the threshold as described above and the other parameters to e.g. ratio = 1:5, attack = 5ms, release = 120ms, knee = 3.5dB, RMS/peak = 0, makeup gain = 0dB.
Level Adjustment:
For global level adjustments you can use Effects -> Amplify (allow clipping, we have a limiter afterwards) or the gain control from the control panel of the audio track.
Peak Limiter:
Use the Fast Lookahead Limiter by Steve Harris (LADSPA plugin), you find it in Effects -> Plugins. Set the limit to -1dB or lower, input gain to 0dB and the release time to e.g. 0.12 seconds.

Another interesting tool is the Levelator. It tries to automate all the discussed steps and does it very nice. So if you just want to throw your audio in and get it back with good levels, use this program.
But be careful, it might produce common artifacts like pumping and also amplifies unwanted sounds like breathing, background noises etc.

Summary

I hope you are convinced now that peak normalization is not the same as loudness normalization and should be avoided.

A detailed loudness post production of podcasts or speech audio should involve the following step:

Loudness normalization: Bring different parts to a uniform volume.
Dynamic range compression: Use a compressor to lower the volume of short and loud spikes.
Global volume adaptation: Raise or lower the volume of the whole audio file to a similar loudness level than other programs.
Peak limiter: Limit all peaks to at least -1dBFS to avoid clipping.

Audio File Formats for Podcasts

grh@auphonic.com (grh) — Wed, 13 Jul 2011 18:27:34 +0200

Many different audio file formats exist for storing recorded audio data on a computer system. This post compares various types and gives suggestions on which format one should use, especially when producing podcasts or other online audio.

If you just want to see some practical tips, skip the description and read the conclusion.

Types of audio formats

There are three main types of audio file formats:

Uncompressed audio formats:
Uncompressed audio formats store the audio information as it is recorded. This results in big files, but no information is lost, therefore they are suitable for archiving original recordings. The most common uncompressed audio format is PCM, which is usually stored in a WAV or AIFF file.
Lossless compression:
Lossless audio compression formats need less space then uncompressed formats, without any loss in quality. They work similar to ZIP, but the compression algorithms are specifically designed for audio data. Some example formats are FLAC (Free Lossless Audio Codec) or ALAC (Apple Lossless Audio Codec), for a comparison of various codecs see Lossless comparison.
Lossly compression:
Lossy compression formats significantly reduce the file size, by throwing away information imperceptible to humans. This gives very small files, but some information is lost and cannot be reconstructed. The best-known example is MP3, others are Ogg Vorbis, AAC, WMA, ...

A great source of information about audio codecs and listening tests is the hydrogenaudio.org wiki and forum.

Relevant audio formats for podcasts

When distributing a podcast or other audio over the internet, you want to have the smallest possible filesize, the best possible quality and everyone should be able to play it (on all operating systems, on mobile phones, portable audio players, car audio players etc.).
Because of the much smaller filesize, lossy formats are the only real option. Additionally one may archive the produced podcast in an uncompressed or lossless compressed audio file.

The characteristics of a few important formats are listed below.

MP3: MPEG-1 Audio Layer 3 (description)

the most widespread and the de facto standard of online audio file formats
most, if not all, hardware players support MP3
acceptable quality, but already quite old
patented: a license is required to "distribute and/or sell decoders and/or encoders" (see e.g. MP3 licensing and patent issues)

Ogg Vorbis: (description, official page)

an open source and patent free audio codec
performs very well from low to high bitrates (more advanced than MP3)
widespread in the open source community
some implementations are more computationally intensive than MP3
some portable players support Ogg Vorbis out of the box (see Vorbis Portable Players)

AAC: Advanced Audio Codec (description)

the latest industry standard and the official successor to MP3 (also called MP4 audio, most common file extension is M4A)
different types of AAC exist (LC AAC, HE AAC, HE AAC v2)
the AAC format is on par with Ogg Vorbis and other modern codecs, HE AAC should provide higher quality at low bitrates
HE AAC uses spectral band replication (high frequencies are removed and calculated from lower frequencies, see SBR), HE AAC v2 adds a parametric stereo method (stereo audio is created out of a mono signal, see PS) - don't use HE AAC for higher bitrates!
heavily patented: a patent license is required for all manufacturers or developers of AAC codecs
widespread in the apple community (iTunes, iPod, iPhones etc.), many other portable players don't support AAC

FLAC: Free Lossless Audio Codec (description, official page)

lossless compression, no information is lost
flac files are typically reduced to 40-60% of their original size
very fast encoding and decoding
open source and patent free

Conclusion

If you want to reach everyone and maintain only one file format, you have only one choice: MP3. It is not the most advanced codec, but everything supports it and the quality is sufficient.

Additionally it may be advantageous to offer your audio in Ogg Vorbis or AAC. Ogg Vorbis is widespread in the open source and creative commons movement and should be used when patent problems are important for you. AAC is the new standard on all "i"-Devices (Apple) and shows very good performance especially at low bitrates (HE AAC).

Finally, if you also want to archive an original version of your audio, use FLAC or just PCM (WAV, AIFF).

Further reading: Podcast Comparison, Part 1: File Formats and Bitrates.

Auphonic at the Science Park Graz Incubator

grh@auphonic.com (grh) — Wed, 29 Jun 2011 17:10:53 +0200

Some weeks ago I presented auphonic to the advisory board of the Science Park Graz, the local version of the Austrian incubator network AplusB.
Fortunately I am happy to announce, that the jury enjoyed the project and therefore auphonic is now part of the Science Park.

What are incubators?

AplusB (Academia plus Business) is the Austrian hightech incubator network. These incubators are linked to academic institutions, in my case the universities in Graz, and they support the development of technology-based projects for a period of 1.5 years.

Here a more detailed (and little bit cryptic) self-definition from the AplusB website:

AplusB - Hightech Incubators in Austria

The Austrian incubators network AplusB (Academia plus Business) is the national and international representation of the eight AplusB centres linked to academic institutions and subsidised by the Federal Ministry for Traffic, Innovation and Technology (FMTIT) under the AplusB program.
The incubators support, accompany and accelerate the development of promising foundation projects by providing counselling, networks, funding and infrastructure supporting the implementation of a business. The founders have access to the entire know-how available through the incubators’ network as well as the networks of the nine incubators. The incubators’ network acts as a representative body for AplusB start-up companies and supports them in specific ways in the fields of growth, funding and internationalisation.
In the circles of partners and networks of AplusB centres, nearly all Austrian universities are represented, as well as universities of applied sciences, research institutions, funding agencies as well as private firms - adding up to over 150 partners.
Until April 2010, 352 foundation projects were accepted into the AplusB program. 289 of them became newly founded companies, 247 of them successfully completed the program. About 250 patents were used and over 1400 jobs were created with an academics rate of 72%. Especially the academic start-ups show above-average growth potential as well as a high intensity of input knowledge. In total more than 150 million Euros in capital were made available to the companies directly.

What does that mean?

First, I finally finished the time-consuming process of writing a business plan :). It took really more time as I thought, but now I have at least something which can be presented to others, let's see if I will ever need it again. Many thanks to the friendly and patient people from Science Park, for correcting various draft versions.

Furthermore, I will get a free office at Science Park, some small financial aids, support in non-technical areas and a close collaboration with scientific partners here in Graz.
I am also able to pay another developer a little bit, if someone wants to join the project.

Coming back to Graz after living in Berlin and Vienna, I am looking forward to some inspiring month working on auphonic and maybe some fruitful collaborations with the various other interesting projects going on at the Science Park.

Django Deployment with Nginx, uWSGI, virtualenv and Fabric

grh@auphonic.com (grh) — Sat, 18 Jun 2011 11:24:33 +0200

As a Python or Django developer, various WSGI and webservers are available to deploy an application. A very fast, lightweight and relatively easy-to-install solution is the combination of Nginx, the high-performance HTTP server, and uWSGI (see benchmark of python WSGI servers). Furthermore, virtualenv and Fabric are invaluable tools to handle Python dependencies and to automate deployments.

This tutorial describes how to setup a Django 1.3, Nginx and uWSGI stack. My OS at the time of this writing was Ubuntu 10.04 LTS.

Installing all requirements

First install all the handy python tools, if you don't have them already (maybe in your virtualenv):

apt-get install python-virtualenv python-pip fabric

Nginx is not yet in the standard Ubuntu packages, therefore one has to add an additional apt repository:

apt-get install python-software-properties
add-apt-repository ppa:nginx/development
apt-get update
apt-get install nginx

To install uWSGI, activate your virtualenv and install it with pip:

source /path/to/your/virtualenv/bin/activate
pip install uwsgi

Configure uWSGI

First you need a WSGI-script for your django project. Put the following file in your django project folder (same folder where settings.py lives) and name it wsgi.py:

import os
import sys
import site

site.addsitedir('/path/to/your/virtualenv/lib/python2.6/site-packages')

os.environ['DJANGO_SETTINGS_MODULE'] = 'settings'

import django.core.handlers.wsgi
application = django.core.handlers.wsgi.WSGIHandler()

Then we need an init script for the uWSGI daemon, I used upstart, because I am lazy. The following file should be in /etc/init/uwsgi.conf, for a description of the uWSGI parameters see uWSGI documentation:

# file: /etc/init/uwsgi.conf
description "uWSGI starter"

start on (local-filesystems and runlevel [2345])
stop on runlevel [016]

respawn

# home - is the path to our virtualenv directory
# pythonpath - the path to our django application
# module - the wsgi handler python script

exec /path/to/your/virtualenv/bin/uwsgi \
--uid your-user \
--home /path/to/your/virtualenv \
--pythonpath /path/to/your/django/project \
--socket /tmp/uwsgi.sock \
--chmod-socket \
--module wsgi \
--logdate \
--optimize 2 \
--processes 2 \
--master \
--logto /path/to/your/log/uwsgi.log

Be careful to specify the right user (must have all necessary permissions, also for the database) - and no, root is not a good idea.

Configure Nginx

Put the Nginx configuration file for your domain in /etc/nginx/sites-available/yourdomain.com:

# file: /etc/nginx/sites-available/yourdomain.com
# nginx configuration for yourdomain.com

server {
        server_name www.yourdomain.com;
        rewrite ^(.*) http://yourdomain.com$1 permanent;
}

server {
        server_name yourdomain.com;
        access_log /path/to/your/log/access.log;
        error_log /path/to/your/log/error.log;

        location / {
                uwsgi_pass unix:/tmp/uwsgi.sock;
                include /etc/nginx/uwsgi_params;
        }

        location /static {
                root /root/path/of/your/static/files;
        }
        location /media {
                root /root/path/of/your/media/files;
        }
}

Then activate your domain:

ln -s /etc/nginx/sites-available/yourdomain.com /etc/nginx/sites-enabled/yourdomain.com

Nginx already ships with an init script, so we don't have to write another one.

Starting and Restarting the Server

To start the server type:

start uwsgi
/etc/init.d/nginx start

Or if you want to restart it:

restart uwsgi
/etc/init.d/nginx restart

And don't forget to set your domain in Django Admin -> Sites!

Deployment with Fabric

Put the following file into /path/to/your/django/project/fabfile.py:

""" Deployment of your django project.
"""

from fabric.api import *

env.hosts = ['yourdomain.com']
env.user = "your-user"

def update_django_project():
    """ Updates the remote django project.
    """
    with cd('/path/to/your/django/project'):
        run('git pull')
        with prefix('source /path/to/your/virtualenv/bin/activate'):
            run('pip install -r your_pip_requirement_file.txt')
            run('python manage.py syncdb')
            run('python manage.py migrate') # if you use south
            run('python manage.py collectstatic --noinput')

def restart_webserver():
    """ Restarts remote nginx and uwsgi.
    """
    sudo("service uwsgi restart")
    sudo("/etc/init.d/nginx restart")

def deploy():
    """ Deploy Django Project.
    """
    update_django_project()
    restart_webserver()

Then you can deploy your project with the command:

fab deploy

Further Resources

The following links helped me to get the system running:

Deploying django projects by Jiri Barton
Deploying your Django applications with Nginx and uWSGI by Martin Rusev
Sexy: NGiNX + uWSGI deployment for Django with virtualenv by Many Nights