Hacker News Reader

Comment by jancsika | Hacker News Reader

Perhaps you could get it smaller by making a "vector format" for podcasts and leave it to the audio player to "do it live"

View on HN

This reminds me of MIDI files and how depending on one's sound device, would sound subtly different.

My Sansa MP3 player, Acer laptop, and gateway PC all had different ways of 'rendering' the Orchestra hit instrument in particular, which made for a rousing Termina Field listening experience on the bus or at Grandma's compared to the family computer.

The last time I looked at the Web Audio spec it appeared to leave rendering to the implementation. So regarding aliasing with OscillatorNode[1]:

"It is expected that an implementation will take some care in achieving this ideal, but it is reasonable to consider lower-quality, less-costly approaches on lower-end hardware."

Luckily nearly no one seems to be using this API so I doubt there really are anywhere near as many "approaches" as you can hear from MIDI file players.

[1] It is expected that an implementation will take some care in achieving this ideal, but it is reasonable to consider lower-quality, less-costly approaches on lower-end hardware.

Style sheets for sound - I think BBC radio used to sell CDs for this. Now they can provide Audio Style Sheets with the acronym of... oh, wait

Golly - you could make any story sound like it was a Radio 4 play

Acronym already taken by Advanced SubStation Alpha subtitle files!

There actually was work on CSS additions to give hints to Text-To-Speech engines, e.g. https://www.w3.org/TR/2018/NOTE-css3-speech-20180605/

I was thinking more that one would use mechanical turk to annotate each soundfile with these tags for the set of all podcasts. Then use that as training data to a ML algorithm that gets used to render the output for each tag.

So I'd use my archiveaudio tag from the example and the algo would output the band-limited frequencies of what sounds vaguely like a cop with speech patterns signaling defensiveness and irritation.

It would all sound perfectly like a podcast if you were listening to it in a sleep state. Then if you concentrated on it you'd realize there's no coherent plot or even any actual natural language being spoken in the entire recording.

We really need some kind of highly compressible alternative format for podcasts. Representing speech with a very limited set of code points. It might be a bit lossy, so you could lose the tone of the voice and other speech characteristics, but there could be some ways to still represent that by carefully picking the words you utter in the podcast.

Of course I'm only spitballing here, this would require some pretty advanced research to be made.

> We really need some kind of highly compressible alternative format for podcasts.

WebVTT, reproduced at the listener side by TTS? At some point, there's a quality chasm that can only be leaped by using a format other than sampled audio.

My question for you is: Why do we need that?

Uh.. isn’t that pretty much what things like lpcnet and wavenet and codec2 are (discussed here) doing? These are codecs that are speech specific. I don’t understand what unique idea you are trying to “spitball” here.

Lots of advanced research has been done in this area. These are the results as of now.

Things will hopefully continue to improve.

If that is all you are after, then the most efficient codec of all is compressed text, and the player a text-to-speech engine.

> We really need some kind of highly compressible alternative format for podcasts.

Do we though? In the current state of technology podcasts can be quite long and still consume less bandwidth than a HD movie trailer.

Might be easier to create 'players' for various peoples' synthesized voices, and just compress and stream the text.

Would be even more interesting if each sound effect were randomized a bit so you don’t feel like you are hearing the same thing over and over.

randomize=”12%”

See also module music. MIDI + SoundFonts would be a reasonable idea for music and sound effects, but it wouldn't be suitable for voice.

Something like Vocaloid parameters would be interesting, but the synthesis engine is probably pretty intensive.

(To be clear, I think this is an idea that's fun to think about as a novelty but probably a dead end otherwise.)

It would hardly surprise me if an addition like this was included in the ridiculous web audio standards

Reply on news.ycombinator.com