I have several tapes (yes actual cassette tapes) of my grandfather reading a novel.
Unfortunately a few of the tapes have degraded to the point that I cannot play them back.
I would love to recreate his voice, to “rerecord” the missing bits.
The recordings are in Danish.
Is this possible?
If it is, how can I go about it?
Clone his vocal chords and get surgery to replace your old voice chords with the new cloned one.
Elvenlabs is currently the best but you can get some very good results with first xtts then rvc as a second pass. It involves fine tuning models and running things with python and notebooks, so requires some know how.
You can explore more models on the huggingface page https://huggingface.co/models?pipeline_tag=text-to-speech&sort=trending
Most have a huggingface space dedicated to them where you can try them, here is the xtts space for example https://huggingface.co/spaces/coqui/xtts
The language adds an other layer of difficulty, I would try their demo first to see if it gives anything workable but it isn’t a language current tts software cater too, it doesn’t seem to be an available option on xtts sadly.
Thank you for the tips. As I see it currently, I expect the language to be the biggest hurdle. It doesn’t appear like something I can add myself, even if I had the data for a model. So as far as I can tell it involves two currently more or less impossible steps: Get model data and teach language to model.
If you have material with him speaking in English, you might be able to train an xtts model on it and then use that to bypass the elvenlabs captcha but I’m not sure if they give enough time. Although GPU rental is cheap these days, so captcha time is less of a factor.
If anything, the tech is moving quite fast, it will definitely be easier in a few years, maybe even months.
If you can get them into a digital format I’ve personally used eleven labs to clone voices and make narrations for missions I created for a video game. I tried using different open source projects and getting it to run on my own with no avail, but 11 labs has been solid (it is unfortunately paid software of like $5/10 bucks a month though)
The very first thing you should do is get them professionally digitized, that way the quality won’t degrade any further. Then you can try training a voice AI, but as long as you have the digitized version, you can always train whatever new AI is invented in the future.
I can’t speak to the AI voice generation part of this, but you might be interested in the Domesday Duplicator for digitizing your audio, especially if some or it is slightly degraded.
https://github.com/harrypm/DomesdayDuplicator
The project was originally designed for laserdisc, but it’s been expanded to support VHS and cassette tape. Traditionally, you would play your tape on a cassette player, then the built in analog circuitry would convert the magnetic signals into audio, amplify them, and feed them to a sound card on your PC, which then converts the analog signal to a digital audio stream.
With the Domesdsy Duplicator, you record the raw magnetic signal from the read head and directly digitize it into a bitstream that you can then process as needed. For DIY archiving from an analog source, it’s one of the best options for signal fidelity, and it will give you the truest representation of what’s actually on the tape.
