Reviving the 1973 Unix text to voice translator | Hacker News Reader

Reviving the 1973 Unix text to voice translator

spinellis.gr

55 pointsingve19 comments

I wonder what it would sound like?

Radio Shack used to sell the SPO256 text to speech chip. I had a bunch of them and the speech was barely understandable. It was worse than Talking Sam on the C-64.

There’s a recording of what the SPO256 sounded like on its Wikipedia page:

https://en.wikipedia.org/wiki/General_Instrument_SP0256

Thanks for posting the wikipedia page with the recording. It brought back memories!

Yes, it is certainly difficult to understand the sound. However, all these years later, the same could be said of remote video connections in which speech is sometimes replaced with undecipherable squeaks and echos.

I'm surprised that zoom et al. have not developed a scheme to show text during these squeaking times. After all, any meeting has people and computers that could detect a temporary problem, mandating a temporary switch to a dictation mode.

The author has uploaded a demo on YT: https://www.youtube.com/watch?v=RO00-eby6_E

There's this Bell Labs video where Lorinda Cherry demonstrates it (shortly after 15:30): https://youtu.be/XvDZLjaCJuw

URL with time: https://youtu.be/XvDZLjaCJuw?t=930

In the same video, there's a fun moment where they demonstrate constructing a simple spell checker out of unix primitives, and the demonstrator comments that "unix" is a word unlikely to ever appear in the dictionary: https://youtu.be/XvDZLjaCJuw?t=551

Of course, now it's in every dictionary!

That’s more like text to gibberish.

The example sounds almost scouser?

Radio Shack is responsible for this chip being commonly known as the SPO256, which was in fact a typo in the documentation for the Radio Shack "Narrator Speech Processor". Its correct name is SP0256, with a zero, not a letter O.

Writing a post about an old program that produces speech and then not adding a sample mp3 is a special kind of cruel.

Text to speech.... several years ago I was playing with a pet project to make procedurally generated number station broadcasts. For that I needed text 2 speech with ?euphony? The available tools were horrid. There was some library MaryTTS (?) that had a poor python wrapper to talk to it. All the web based systems were super stingy in their demo resources and their use cases seemed geared more for TTS phrases that you’d have on a phone menu or something, not a paragraph of prose. I used Watson’s tool for a while before they became very annoying and I abandoned it.

Is there any better TTS Tool people could recommend?

Amazon Polly [0] and Google's Neural Voices [1] make Watson's voice seem pretty awful by comparison. Unfortunately, they can be costly, depending on what you're doing, and you have to hope neither of those companies cut off your access for arbitrary reasons.

If/when that happens, and you go to look at the rest of the tools... They're pretty crap. Still.

[0] https://aws.amazon.com/polly/

[1] https://cloud.google.com/text-to-speech/

I use Azure neural TTS for (Chinese) voice synthesis in my app, and honestly it's amazing.

I did experiment with Google's API at the beginning, the audio would contain recurring artefacts that made it painfully obvious it wasn't real. If I had to describe it, I'd say "50ms of the sound your computer makes when it hangs while playing an audio file".

No such problems with Azure. Zero artefacts and nice, crisp, natural enunciation, though unfortunately it is still quite unreliable when synthesizing single words (as compared with full sentences). You can actually test it out by using the "Read Aloud" feature inside Edge.

I don't know how well this translates to English, though. I definitely enjoy the benefit of working with a gender/language combination (female/Chinese) that's received the largest total investment of effort and resources. Many well-resourced Chinese companies (banks, tech services, gaming, etc) all use similar synthesized female voices in their phone/online/B&M services.

If your software works on a Mac, maybe the built-in TTS functionality would work.

The term I meant to use above was 'prosody' not euphony.

I've been working on a TTS API connector and SSML editor. Voxabot.com Take note that unfortunately the Neutral TTS for Polly isn't fully exposed. That’s arguably the best sounding TTS.

I can't access the demo. It says "Sign in with Google".

What does emacspeak use for example?

For open source offline TTS with more or less recent algorithms you can check

https://github.com/TensorSpeech/TensorFlowTTS

It should be possible to run the results of Unix speak.c through a real (or emulated) Votrax Type 'n Talk unit. I know MAME has emulation of said unit, and while it isn't perfect, its pretty understandable.

I took a shot at making a modernized port of speak.c myself not long after the code was found, but it didn't get very far, sadly. I couldn't figure out an easy way of dealing with the multiple-character character constants.

Reply on news.ycombinator.com