chipmusic.org — speech synthesis thread (2012)

Someone asks whether a three-to-four-minute spoken-word piece is possible with lsdj's SPEECH instrument. The answers are the only first-hand account of the speech instrument in this wiki, and they are discouraging in a useful way.

Key claims#

• What it costs. Jotie, who has actually done it — "Mortal Kombat", "This one is dedicated to all the ravers of the nation", "kill three million people and become god", "resistance is futile", "expialidocious" — calls it "slow, difficult and in my opinion, flawed. You can't delete wrong letters." On the spoken-word idea: "I don't think it's possible to generate enough words with 4F and speech kit instruments combined to achieve what you want. Maybe a poem, but I don't expect much room for a song." Zef agrees: "I've gotten decent at making words legible, but it takes pretty much forever."
• The workflow trick that makes it tractable (XyNo): "Download an Atari ST emulator and write your lines in STSPEECH — it will show you how to do what you just wrote in allophones." The Atari's speech program does the English-to-allophone translation, and you copy the result into LSDj's word screen. Steem SSE is the suggested emulator; another poster calls it "an amazing find. I will be adding more speech now."
• Mixing is what makes speech intelligible (Jansaw): "mixing is the only thing that is going to make your speech stand out. Panning certain tracks left and right, while keeping the wav (speech) centre will help it come through the mix … Give the listeners a soundscape." And phrasing: "humanising the speech — don't put them on every downbeat tick, try and give them space and change the syllable when it's natural to do so."
• A pitch-shift trick on the speech synth is demonstrated in a linked video, and a save attributed to roboctopus is described as CHK R02 DIZ R02 OUT R02 — "which sounds really cool".
• The original poster's own conclusion is a workflow one: a long piece would need many .sav files and a DAW to stitch the sentences together.

Practical takeaways#

• Write the allophones somewhere else (an Atari ST speech program) and transcribe them; do not compose them in the word screen, where mistakes cannot be deleted.
• Speech lives in the centre of the mix with everything else panned away from it.
• Space the syllables musically rather than on every tick — the phrasing does as much for intelligibility as the allophones.
• Scale the ambition to the tool: a phrase or a hook, not a monologue.

Notable quotes#

"It's slow, difficult and in my opinion, flawed. You can't delete wrong letters." — Jotie

"Mixing is the only thing that is going to make your speech stand out." — Jansaw

Relevance to the wiki#

lsdj documents the SPEECH instrument's mechanics from the manual — 14 word slots built from 59 allophones — but nothing about using it. This is the practice: the cost, the outside tool, and the mixing that makes it land.

To verify#

• The 4F referred to as a limit is presumably the word-slot/allophone capacity, not stated plainly in the thread (unverified).
• The CHK R02 DIZ R02 OUT R02 save is attributed to roboctopus with an "I think" (unverified).

Pages touched#

lsdj · stereo-panning · instrument-design · chipmusic-org · roboctopus