Personally I don’t know of any specific research but afaik the reason why our voices sound terrible to us when recorded is because they’re missing the bass that’s transmitted through the skull, i.e. they sound higher pitched on recording. So I’d try increasing the amplitude of the lower frequencies on the EQ, it’ll take a bit of fiddling to get the right balance.

I’ve found in my experience that over the years my internal voice updated to match how I sound when recorded. So when I hear myself speaking on a recording it’s much less jarring now since it feels much more like how I predicted it would sound.

I know this doesn’t help your current situation but it’s a fun fact I recognized since you mentioned it.

A very simple approximation of your voice as it’s heard if you’re facing someone when speaking would be using a unidirectional mic and recording yourself with the mic pointing the opposite direction as it normally would be (in other words— with the polarity reversed).

A slightly better approximation would be if you did the same thing but with two unidirectional mics pointed at slight angles (with the polarity still reversed) to simulate the placement of your ears.

Obviously the quality of the mic would factor in as well—you’d want mics with a flat frequency curve. To get even pickier you’d also want to use headphones or speakers with a flat frequency response to listen to it. Once you had the recording you could even take impulse responses of certain rooms and process the audio to get an idea of how you sound to others in specific rooms!

deleted by creator

Please God no. I cannot fathom how people manage to hear my real voice.

I sound so much better inside my skull

So to clarify, I am really curious about speech comprehension if a Text To Speech AI is trained to sound like ourselves (myself).

Over the last couple of weeks I have been working on learning about and using offline AI. I don’t know why, but I decided to ask here first.

If anyone is curious, this is what my current favorite LLM has to say (Llama 2 70B Q4 GGML):

You

what does it take to tune a recording of my voice and match it to how I hear my own voice?

Assistant

To tune your recorded voice so that it matches the way you perceive your own voice, there are several steps involved in audio processing. Here’s an overview of the process:

• Recording: Firstly, ensure that you have a high-quality recording of your voice using proper equipment such as a microphone and soundproofing to minimize background noise. This will provide a good starting point for further editing.

• Equalization (EQ): Use an equalizer to adjust the frequency balance in your recorded audio. You can boost or cut specific frequencies that are too prominent or lacking, respectively. For example, if you feel like your voice sounds nasally, you may want to reduce some of the mid-range frequencies around 1kHz and above.

• Compression: Apply compression to even out the dynamic range in your recording. This will help balance the loudness between different parts of your speech or singing, making it sound more consistent.

• Reverb/Delay: Add a small amount of reverb or delay to create a sense of space and depth around your voice. This can make it feel like you’re hearing yourself in a room rather than through headphones or speakers.

• Stereo Imaging: Use stereo imaging tools to adjust the width and positioning of your voice within the mix. This will help create an illusion that your voice is coming from a specific location, making it sound more natural.

• Monitoring: Finally, listen back to your processed audio on different devices such as headphones or speakers in various environments (e.g., quiet room vs noisy cafe) to ensure that the changes you’ve made translate well across different listening scenarios. Make further adjustments if necessary until you are satisfied with how it sounds.

Remember, everyone perceives their own voice differently due to bone conduction and other factors, so there may be some trial-and-error involved in finding a sound that feels right for you. It’s also important to keep in mind that the goal is not necessarily to make your recorded voice identical to how it sounds when speaking but rather to create an enjoyable listening experience for others who will hear it through speakers or headphones.

You

Is there any formulaic approximation or rule regarding bone conduction and how humans perceive their own voice?

Assistant

Bone conduction is the process by which sound vibrations are transmitted directly to the inner ear via the skull, bypassing the outer and middle ears. This occurs when we speak or sing, causing us to hear our voices differently than others do because of this direct transmission pathway.

There isn’t a specific formulaic approximation or rule that can accurately predict how each individual perceives their own voice through bone conduction since it varies from person to person due to factors such as head shape and size, vocal tract resonance, and the acoustics of the surrounding environment. However, in general, people tend to hear their voices as deeper or richer when speaking because low-frequency sounds are more efficiently transmitted through bone conduction than high frequencies.

It’s important to note that while understanding how we perceive our own voice can be helpful for tuning a recording of it, the goal should ultimately be to create an enjoyable listening experience for others who will hear your recorded voice through speakers or headphones rather than trying to perfectly replicate what you hear when speaking.

My guess is there’s some subjective components to it that simple changes wouldn’t be able to adjust to, each person will have to do the changes themselves to make it match what they’re hearing…

I always wondered if this is possible.

I’m not convinced it’s possible. My internal voice has a different accent, not just different acoustic properties.

I’ve been thinking about this exact same thing just today. Spooky.