For my comprehensive channel trailer, I created a word cloud of the words used in titles and descriptions of the videos uploaded each month. Word clouds have been around for a while now, so that’s nothing unusual. For the soundtrack, I wanted to make audio versions of these word clouds using text-to-speech, with the most common words being spoken louder. This way people with either hearing or vision impairments would have a somewhat similar experience of the trailer, and people with no such impairments would have the same surplus of information blasted at them in two ways.

I checked to see if anyone had made audio word clouds before, and found Audio Cloud: Creation and Rendering, which makes me wonder if I should write an academic paper about my audio word clouds. That paper describes an audio word cloud created from audio recordings using speech-to-text, while I wanted to create one from text using text-to-speech. I was mainly interested in any insights into the number of words we could perceive at once at various volumes or voices. In the end, I just tried a few things and used my own perception and that of a few friends to decide what worked. Did it work? You tell me.

Voices

There’s a huge variety of English voices available on macOS, with accents from Australia, India, Ireland, Scotland, South Africa, the United Kingdom, and the United States, and I’ve installed most of them. I excluded the voices whose speaking speed can’t be changed, such as Good News, and a few novelty voices, such as Bubbles, which aren’t comprehensible enough when there’s a lot of noise from other voices. I ended up with 30 usable voices. I increased the volume of a few which were harder to understand when quiet.

I wondered whether it might work best with only one or a few voices or accents in each cloud, analogous to the single font in each visual word cloud. That way people would have a little time to adapt to understand those specific voices rather than struggling with an unfamiliar voice or accent with each word. On the other hand, maybe it would be better to have as many voices as possible in each word cloud so that people could distinguish between words spoken simultaneously by voice, just as we do in real life. In the end I chose the voice for each word randomly, and never got around to trying the fewer-distinct-voices version. Being already familiar with many of these voices, I’m not sure I would have been a good judge of whether that made it easier to get used to them.

Arranging the words

It turns out making an audio word cloud is simpler than making a visual one. There’s only one dimension in an audio word cloud — time. Volume could be thought of as sort of a second dimension, as my code would search through the time span for a free rectangle of the right duration with enough free volume. I later wrote an AppleScript to create ‘visual audio word clouds’ in OmniGraffle showing how the words fit into a time/volume rectangle.  I’ve thus illustrated this post with a visual word cloud of this post, and a few audio word clouds and visual audio word clouds of this post with various settings.

However, words in an audio word cloud can’t be oriented vertically as they can in a visual word cloud, nor can there really be ‘vertical’ space between two words, so it was only necessary to search along one dimension for a suitable space. I limited the word clouds to five seconds, and discarded any words that wouldn’t fit in that time, since it’s a lot easier to display 301032 words somewhat understandably in nine minutes than it is to speak them. I used the most common (and therefore louder) words first, sorted by length, and stopped filling the audio word cloud once I reached a word that would no longer fit. It would sometimes still be possible to fit a shorter, less common word in that cloud, but I didn’t want to include words much less common than the words I had to exclude.

I set a preferred volume for each word based on its frequency (with a given minimum and maximum volume so I wouldn’t end up with a hundred extremely quiet words spoken at once) and decided on a maximum total volume allowed at any given point. I didn’t particularly take into account the logarithmic nature of sound perception. I then found a time in the word cloud where the word would fit at its preferred volume when spoken by the randomly-chosen voice. If it didn’t fit, I would see if there was room to put it at a lower volume. If not, I’d look for places it could fit by increasing the speaking speed (up to a given maximum) and if there was still nowhere, I’d increase the speaking speed and decrease the volume at once. I’d prioritise reducing the volume over increasing the speed, to keep it understandable to people not used to VoiceOver-level speaking speeds. Because of the one-and-a-bit dimensionality of the audio word cloud, it was easy to determine how much to decrease the volume and/or increase the speed to fill any gap exactly. However, I was still left with gaps too short to fit any word at an understandable speed, and slivers of remaining volume smaller than my per-word minimum.

I experimented with different minimum and maximum word volumes, and maximum total volumes, which all affected how many voices might speak at once (the ‘hubbub level’, as I call it). Quite late in the game, I realised I could have some voices in the right ear and some in the left, which makes it easier to distinguish them. In theory, each word could be coming from a random location around the listener, but I kept to left and right — in fact, I generated separate left and right tracks and adjusted the panning in Final Cut Pro. Rather than changing the logic to have two separate channels to search for audio space in, I simply made my app alternate between left and right when creating the final tracks. By doing this, I could increase the total hubbub level while keeping many of the words understandable. However, the longer it went on for, the more taxing it was to listen to, so I decided to keep the hubbub level fairly low.

The algorithm is deterministic, but since voices are chosen randomly, and different voices take different amounts of time to speak the same words even at the same number of words per minute, the audio word clouds created from the same text can differ considerably. Once I’d decided on the hubbub level, I got my app to create a random one for each month, then regenerated any where I thought certain words were too difficult to understand.

Capitalisation

In my visual word clouds, I kept the algorithm case-sensitive, so that a word with the same spelling but different capitalisation would be counted as a separate word, and displayed twice. There are arguments for keeping it like this, and arguments to collapse capitalisations into the same word — but which capitalisation of it? My main reason for keeping the case-sensitivity was so that the word cloud of Joey singing the entries to our MathsJam Competition Competition competition would have the word ‘competition’ in it twice.

Sometimes these really are separate words with different meanings (e.g. US and us, apple and Apple, polish and Polish, together and ToGetHer) and sometimes they’re not. Sometimes these two words with different meanings are pronounced the same way, other times they’re not. But at least in a visual word cloud, the viewer always has a way of understanding why the same word appears twice. For the audio word cloud, I decided to treat different capitalisations as the same word, but as I’ve mentioned, capitalisation does matter in the pronunciation, so I needed to be careful about which capitalisation of each word to send to the text-to-speech engine. Most voices pronounce ‘JoCo’ (short for Jonathan Coulton, pronounced with the same vowels as ‘go-go’) correctly, but would pronounce ‘joco’ or ‘Joco’ as ‘jocko’, with a different vowel in the first syllable. I ended up counting any words with non-initial capitals (e.g. JoCo, US) as separate words, but treating title-case words (with only the initial letter capitalised) as the same as all-lowercase, and pronouncing them in title-case so I wouldn’t risk mispronouncing names.

Further work

A really smart version of this would get the pronunciation of each word in context (the same way my rhyming dictionary rhyme.science finds rhymes for the different pronunciations of homographs, e.g. bow), group them by how they were pronounced, and make a word cloud of words grouped entirely by pronunciation rather than spelling, so ‘polish’ and ‘Polish’ would appear separately but there would be no danger of, say ‘rain’ and ‘reign’ both appearing in the audio word cloud and sounding like duplicates. However, which words are actually pronounced the same depend on the accent (e.g. whether ‘cot’ and ‘caught’ sound the same) and text normalisation of the voice — you might have noticed that some of the audio word clouds in the trailer have ‘aye-aye’ while others have ‘two’ for the Roman numeral ‘II’.

Similarly, a really smart visual word cloud would use natural language processing to separate out different meanings of homographs (e.g. bow🎀, bow🏹, bow🚢, and bow🙇🏻‍♀️) and display them in some way that made it obvious which was which, e.g. by using different symbols, fonts, styles, colours for different parts of speech. It could also recognise names and keep multi-word names together, count words with the same lemma as the same, and cluster words by semantic similarity, thus putting ‘Zoe Keating’ near ‘cello’, and ‘Zoe Gray’ near ‘Brian Gray’ and far away from ‘Blue’. Perhaps I’ll work on that next.

I’ve recently been updated to a new WordPress editor whose ‘preview’ function gives a ‘page not found’ error, so I’m just going to publish this and hope it looks okay. If you’re here early enough to see that it doesn’t, thanks for being so enthusiastic!