speech-synthesis on S Anand

Things I Learned - 21 Jan 2024

Sun, 21 Jan 2024 00:00:00 +0000

This week, I learned:

When comparing Mistral with 4b quantization vs unquantized:
- 2 responses were significantly shorter and fairly different
- 1 was identical
- 1 was almost identical but shorter by a few words
- 1 was slightly longer and fairly different
#PREDICTION As humans have more conversations with LLMs, they will replace video watching and interactive gaming with conversation based role play. New game genres will evolve
Lilac is an LLM-based data curation tool. Use it to search by concept (e.g. PII, duplicates, etc.) and then drop/update the results.
Lungs have a Hausdorff dimension of 2.97 – giving them one of the highest surface area to volume ratio. Brains are 2.8. Sierpinski Pyramid is exactly 2 – which is weird. To solid-paint twice the size, you need 4 times as much paint.
How I write podcast. Tim Ferriss
- High bars are constraints. I set the strongest constraints against the scarcest resources. Like reputation
- Being a category of one is more defensible than a competitive advantage
- Content always beats presentation. When in doubt, push for more interesting content
- Regular publishing improves thinking
- To build a habit, do less than you think you can do. That makes it easier to build momentum on the habit and sustain during crunch times
- There is a lot of mediocrity in the world. If you’re doing something (in a winner take all ecosystem), be the best.
- Top lawyers are exceptional proofreaders. They are able to see what is unclair, and what is redundant, and what has loop holes very quickly.
- Forcing yourself to cut down from a thousand words to 200 to a paragraph to a sentence takes you through a phase transition where you discover something unexpected
- The more outrageous the question, the more likely it is to be useful in generating a new perspective
Eleven-labs speech synthesis with voice cloning is at the uncanny valley. With two 5-minute samples, my voice sounds a fair bit like my voice but is very clearly not my voice. I find stability ~ 30%, similarity ~ 80% and style ~50% gives a reasonable outcome. But the default voices (e.g. Joseph, George, Charlie) are excellent.
Practical AI podcast: AI predictions for
AI by API is the norm today and will grow
- Just having AI is no longer a differentiator
- AI is part of life, not just work
- #TODO Explore quickdrop from Stability for Maruti
- #TODO Explore Codium VS Code plugin and Continue.dev
- Hybrid systems that combine stats, ML, DL and AI models will grow
- AGI and AutoGPT resurgence
- RAG will continue to be a focus
- GPT4 will be beaten by open source models. Special purpose models beat it already
- Self hosted and cloud hosted models will grow for security
- Small language models will grow
- Productivity will be enhanced rather than replaced
- Multi modal models will grow
- Cost efficiency will grow in focus
GPT Builder help explains how the GPT Builder updates GPTs - including some very interesting prompts

Experiments in sound

Tue, 12 Sep 2006 12:00:00 +0000

Wikipedia says the human voice frequency for speech is between 85 to 155 Hz for men, and 165 to 255 Hz for women. That set me thinking.

What is the limit to our hearing?
How do sounds differ?
How can we synthesise speech?

What are the limits to our hearing?

Kids can hear frequencies from 20 Hz to 20 kHz, while adults hear only up to 12-14 kHz (Frequency Range of Human Hearing).

To check the lower frequency limit, I created an MP3 with sounds from 1 Hz to 100 Hz at 1 second intervals. Just play the sound, and see when you start hearing something. (Of course, whether you can hear something also depends on the volume of your speaker, the ambient noise, etc.) I could hear nothing for the first 40 seconds: so I can’t hear frequencies lower than 40 Hz.

PS: Don’t be worried if you don’t hear anything for a while. You’re not supposed to! Keep the volume at full level, though.

To check the upper frequency limit, I created this MP3 with sounds from 1 kHz to 20 kHz in 1 second intervals. Just play the sound, and see when you stop hearing anything. I couldn’t hear anything beyond 14 seconds: so I can’t hear frequencies beyond 14 kHz.

How do sounds differ?

I took this audio file of someone reciting vowels and plotted a spectrogram (below). A spectrogram plots time on the X axis and frequency on the Y-axis.

Some observations:

All the vowels have evenly spaced bars. (In this case, they’re all multiples of something around 120 Hz.)
‘u’ has the lowest frequency mix. ‘a’ spans from low to high. ‘i’ has a bit of low and a bit of high, nothing in the middle. ‘ai’ and ‘au’ look like ‘a’ followed by ‘i’ and ‘u’ respectively.

How can we synthesise speech?

I don’t know. There are lots of speech synthesizers. They sound robotic. I’m trying to see if knowing what sounds look like improves things. I’ll let you know if I do well.

Comments

Dhar 13 Sep 2006 5:09 am: How did you go about synthesizing the MP3s with increasing frequencies? Did you use some software for that? Similarly what software did you use to plot the spectrogram?
S Anand 13 Sep 2006 6:10 am: I wrote a perl script to generate the frequencies, and Spectrogram v.0.8 to create the spectrograms.
Sumit Dhar 16 Sep 2006 9:23 am: Dude, just downloaded the files. Is the sound stereophonic or monoaural? If it is mono, do you think creating an MP3 with stereophonic sound will make a difference? Cheers, D.
Sumit Dhar 16 Sep 2006 9:25 am: Loved the experiment. Lower: 42 Hz, Higher: 15 KHz
S Anand 16 Sep 2006 9:38 am: It’s mono. I doubt if stereo would make a difference. I’d be playing the same frequency on both speakers, which is what a mono MP3 does anyway.
Dhar 17 Sep 2006 2:05 am: The reason I asked about whether the sound is Mono or Stereo is because of my experience during eye tests. Some letters that were not visible when viewed with just one eye, were clearly readable when using both eyes. Was wondering if the same may be applicable in this case?
S Anand 17 Sep 2006 7:36 am: Ah, I see what you mean. Since both speakers would be playing the sound, hopefully this should not be a factor.

A Glove

Thu, 24 Jan 2002 12:00:00 +0000

I wouldn’t mind a glove.