![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Today's the day! It's happening!
Sep. 24th, 2022 01:22 pm(well, September 21st was the day, but I found out about it this morning)
Automated! Offline! Open-source! Reputable-author! Transcription! That doesn't make you write your own interface software!
(you *can* write your own interface software, if you want, but you don't *have* to)
It even has multiple languages and the option for foreign-speech-to-English-text, though *currently* it sucks at the languages I'm likely to encounter. I'll stick with English-only for now.
Let's feed it the first minute of my standard test audio: me reading aloud the prologue of Ptolemy's Gate. For comparison, here's the real text:
Alexandria: 125 B.C.
The assassins dropped into the palace grounds at midnight, four fleet shadows dark against the wall. The fall was high, the ground was hard; they made no more sound on impact than the pattering of rain. Three seconds they crouched there, low and motionless, sniffing at the air. Then away they stole, through the dark gardens, among the tamarisks and date palms, toward the quarters where the boy lay at rest. A cheetah on a chain stirred in its sleep; far away in the desert, jackals cried.
They went on pointed toe-tips, leaving no trace in the long wet grass. Their robes flittered at their backs, fragmenting their shadows into wisps and traces. What could be seen? Nothing but leaves shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fronds. No sight, no noise. A crocodile djinni, standing sentry at the sacred pool, was undisturbed though they passed within a scale's breadth of his tail. For humans, it wasn't bad--
("it wasn't badly done", but the audio is cut off at the one-minute mark)
[note: running this command consumed 2 CPU threads and about 2.5GB of RAM]
Output:
Alexandria, 125 B.C. The Assassins dropped into the palace grounds at midnight, four fleet shadows dark against the wall. The fall was high, the ground was hard, they made no more sound and impact in the pattern of rain. Three seconds they crouched there, low and motionless sniffing at the air. Then away they stole, through the dark gardens among the town risks and date palms, towards the quarters where the boy lay at rest. A tree on a chain stirred in its sleep. Far away in the desert, jackals cried. They went on pointed toe-tips, leaving no trace in the long, wet grass. Their robes flittered at their backs, fragmenting their shadows in the wisps and traces. What could be seen? Nothing but leaves, shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fronds. No sight, no noise. The crocodile genie, standing sentry at the sacred pool, was on the stirrup, though they passed through the scales, breathless his tail. For humans, it wasn't bad.
*Fuck yes*. Oh, it's not perfect, but it's *basically* intact. And, let's face it, my voice is *not* easy-mode, and neither is this vocabulary: I'm especially impressed that it understood "their robes flittered", when honestly if I didn't already know that's what it was I might not have transcribed it correctly myself.
Very slow, though: closer to six times *slower* than real-time than the claimed six times faster (on default settings). (Probably my hardware is underpowered: I'm running it on a six-year-old laptop aimed at businessfolk.) That is...slow enough that it may actually be unable to keep up with the amount of audio I produce, let alone work on the backlog. (Although, at 2 CPU threads and 2.5 GB of RAM, I could probably run two Whispers in parallel overnight, assigning each of them a different series of files. And I wonder if I could rope in the TV's prosthetic brain...)
Hmm...
Alexander, 125 BC. The assassins dropped into the palace grounds at midnight, four fleet shadows darkened the wall. The fall was high, the ground was hard, they made no more sound on impact in the pattern of rain. Three seconds they crouched there, low and motionless sniffing at the air, then away they stole, through the dark gardens among the cameras and date palms, towards the corners where the boy lay at rest. A cheetah on a chain stirred in sleep, far away and desert jackals cried. They went on pointed toe tips leaving no trace in the long lit grass. The ropes were there at their backs, fragmenting their shadows in the wrists and traces. What could be seen? Nothing but leaves, shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fronds. No sight? No noise. A crocodile, genie, standing sentry at a quick sacred core, will end the story of the past of the skills best with his tail. For humans, it wasn't bad.
My inner 00's selves are impressed to have even made it *that* far, but it's not quite good enough to do actual work with. (Although it's interesting that *this* one correctly got "cheetah" when small.en didn't.)
Alexandria, 125 BC. The assassins dropped into the palace grounds at midnight, four fleet shadows dark against the wall. The fall was high, the ground was hard, they made no more sound on impact in the pattern of rain. Three seconds they crouched there, low and motionless, the thing at the air. Then away they stole, through the dark gardens among the town risks and date poems, towards the quarters where the boy lay at rest. A cheetah on a chain stirred in its sleep, far away and desert jaggles cried. They went on pointed toe tips, leaving no trace in the long wet grass. Their robes swillied at their backs, fragmenting their shadows in the wisps and traces. What could be seen? Nothing but leaves shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fawns. No sight, no noise. The crocodile genie, standing sentry at the sacred pool, was undister of the late past when his scales pressed his tail. For humans, it wasn't bad.
*Slightly* better overall than tiny.en, but not by much, and not strictly superior (it makes some mistakes that tiny.en didn't).
---
If anyone has the hardware specs to pull off running "medium.en" or "large", I'd be interested to see what you get out of putting the clip through it. (Though I can't outsource most of my transcription needs to anyone else, for privacy reasons: this would just be for curiosity's sake, and to know what to look forward to when I someday get my hands on higher-grade hardware myself.)
---
Overall verdict: I am so fucking hyped for this development†, and especially about what can be pulled off if you combine it with Recoll indexing.
The transcription software we (I) have been hoping for is here, right now.
---
†my brain has never been good at generating excitement qualia, but intellectually I am so fucking hyped and I'm not *completely* incapable of emotional excitement ↩
---
(edit: part 2)
Automated! Offline! Open-source! Reputable-author! Transcription! That doesn't make you write your own interface software!
(you *can* write your own interface software, if you want, but you don't *have* to)
It even has multiple languages and the option for foreign-speech-to-English-text, though *currently* it sucks at the languages I'm likely to encounter. I'll stick with English-only for now.
Let's feed it the first minute of my standard test audio: me reading aloud the prologue of Ptolemy's Gate. For comparison, here's the real text:
Alexandria: 125 B.C.
The assassins dropped into the palace grounds at midnight, four fleet shadows dark against the wall. The fall was high, the ground was hard; they made no more sound on impact than the pattering of rain. Three seconds they crouched there, low and motionless, sniffing at the air. Then away they stole, through the dark gardens, among the tamarisks and date palms, toward the quarters where the boy lay at rest. A cheetah on a chain stirred in its sleep; far away in the desert, jackals cried.
They went on pointed toe-tips, leaving no trace in the long wet grass. Their robes flittered at their backs, fragmenting their shadows into wisps and traces. What could be seen? Nothing but leaves shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fronds. No sight, no noise. A crocodile djinni, standing sentry at the sacred pool, was undisturbed though they passed within a scale's breadth of his tail. For humans, it wasn't bad--
("it wasn't badly done", but the audio is cut off at the one-minute mark)
whisper voice-sample-speaking-age-25-1-minute-clip.mp3 --model small.en
[note: running this command consumed 2 CPU threads and about 2.5GB of RAM]
Output:
Alexandria, 125 B.C. The Assassins dropped into the palace grounds at midnight, four fleet shadows dark against the wall. The fall was high, the ground was hard, they made no more sound and impact in the pattern of rain. Three seconds they crouched there, low and motionless sniffing at the air. Then away they stole, through the dark gardens among the town risks and date palms, towards the quarters where the boy lay at rest. A tree on a chain stirred in its sleep. Far away in the desert, jackals cried. They went on pointed toe-tips, leaving no trace in the long, wet grass. Their robes flittered at their backs, fragmenting their shadows in the wisps and traces. What could be seen? Nothing but leaves, shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fronds. No sight, no noise. The crocodile genie, standing sentry at the sacred pool, was on the stirrup, though they passed through the scales, breathless his tail. For humans, it wasn't bad.
*Fuck yes*. Oh, it's not perfect, but it's *basically* intact. And, let's face it, my voice is *not* easy-mode, and neither is this vocabulary: I'm especially impressed that it understood "their robes flittered", when honestly if I didn't already know that's what it was I might not have transcribed it correctly myself.
Very slow, though: closer to six times *slower* than real-time than the claimed six times faster (on default settings). (Probably my hardware is underpowered: I'm running it on a six-year-old laptop aimed at businessfolk.) That is...slow enough that it may actually be unable to keep up with the amount of audio I produce, let alone work on the backlog. (Although, at 2 CPU threads and 2.5 GB of RAM, I could probably run two Whispers in parallel overnight, assigning each of them a different series of files. And I wonder if I could rope in the TV's prosthetic brain...)
Hmm...
whisper voice-sample-speaking-age-25-1-minute-clip.mp3 --model tiny.en
Alexander, 125 BC. The assassins dropped into the palace grounds at midnight, four fleet shadows darkened the wall. The fall was high, the ground was hard, they made no more sound on impact in the pattern of rain. Three seconds they crouched there, low and motionless sniffing at the air, then away they stole, through the dark gardens among the cameras and date palms, towards the corners where the boy lay at rest. A cheetah on a chain stirred in sleep, far away and desert jackals cried. They went on pointed toe tips leaving no trace in the long lit grass. The ropes were there at their backs, fragmenting their shadows in the wrists and traces. What could be seen? Nothing but leaves, shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fronds. No sight? No noise. A crocodile, genie, standing sentry at a quick sacred core, will end the story of the past of the skills best with his tail. For humans, it wasn't bad.
My inner 00's selves are impressed to have even made it *that* far, but it's not quite good enough to do actual work with. (Although it's interesting that *this* one correctly got "cheetah" when small.en didn't.)
whisper voice-sample-speaking-age-25-1-minute-clip.mp3 --model base.en
Alexandria, 125 BC. The assassins dropped into the palace grounds at midnight, four fleet shadows dark against the wall. The fall was high, the ground was hard, they made no more sound on impact in the pattern of rain. Three seconds they crouched there, low and motionless, the thing at the air. Then away they stole, through the dark gardens among the town risks and date poems, towards the quarters where the boy lay at rest. A cheetah on a chain stirred in its sleep, far away and desert jaggles cried. They went on pointed toe tips, leaving no trace in the long wet grass. Their robes swillied at their backs, fragmenting their shadows in the wisps and traces. What could be seen? Nothing but leaves shifting in the breeze. What could be heard? Nothing but the wind sighing among the palm fawns. No sight, no noise. The crocodile genie, standing sentry at the sacred pool, was undister of the late past when his scales pressed his tail. For humans, it wasn't bad.
*Slightly* better overall than tiny.en, but not by much, and not strictly superior (it makes some mistakes that tiny.en didn't).
---
If anyone has the hardware specs to pull off running "medium.en" or "large", I'd be interested to see what you get out of putting the clip through it. (Though I can't outsource most of my transcription needs to anyone else, for privacy reasons: this would just be for curiosity's sake, and to know what to look forward to when I someday get my hands on higher-grade hardware myself.)
---
Overall verdict: I am so fucking hyped for this development†, and especially about what can be pulled off if you combine it with Recoll indexing.
The transcription software we (I) have been hoping for is here, right now.
---
†my brain has never been good at generating excitement qualia, but intellectually I am so fucking hyped and I'm not *completely* incapable of emotional excitement ↩
---
(edit: part 2)
