voice-cloning on S Anand

MGR via ElevenLabs

Sun, 29 Mar 2026 22:31:33 +0800

I was watching Vaa Vaathiyar which has a short clip of MGR speaking. It’s either AI-generated or mimic-ed and it wasn’t bad.

I used ffmpeg to record the audio from the film, transcribed it via Gemini 3 Pro on AI Studio with the prompt:

Transcribe this into Tamil

… which gave me:

ராமு… என்ன செய்திருக்கிறாய் நீ… வாத்தியார் கேட்கிறேன் சொல் நிமிர்ந்து பார்க்க கூட தைரியம் இல்லையா… ஓடாதே… நில்…

Translation:

Ramu… What have you done… Vaathiyar (MGR) is asking, tell me Don’t you have the courage to stand up and look at me… Don’t run… stop…

(GitHub Copilot’s auto-complete translated the above for me as I typed - flawlessly. It’s getting better by the day!)

Then, I used yt-dlp to download the audio from this MGR Short Clip.

Here’s the sample:

I fed this into ElevenLabs’ Instant Voice Clone that needs just 10 seconds of audio and created an “MGR” voice.

Here’s the same dialogue in the cloned voice:

Personally, I think the ElevenLabs version is slightly better. Of course, given the pace of AI improvement, this might just be the impact of a new model release.

2 Apr 2026: Here’s the non-cloned generation from Sarvam’s text to speech with Bulbul v3 standard quality. It feels pretty weak.

Gemini 2.5 Pro Preview TTS gave me this, which feels much better.

Voice coding is the new live coding

Sun, 21 Sep 2025 11:18:17 +0000

In Feb 2025 at PyConf Hyderabad, I tried a new slide format: command-line slideshows in bash.

I’ve used this format in more talks since then:

LLMs in the CLI, PyCon Singapore, Jun 2025
Agents in the CLI, Singapore Python User Group, Jul 2025
DuckDB is the new Pandas, PyCon India, Sep 2025

It’s my favorite format. I can demo code without breaking the presentation flow.
It also draws interest. My setup was the top question in my PyCon talk.

In Sep 2025, at PyCon India, I extended the setup for voice typing. talkcode.sh is a Bash pipeline that:

uses ffmpeg to record mic into as 16 kHz mono, voice-optimized .opus
sends audio to Gemini via llm for transcription
uses awk to extract the code fence
uses xclip to copy the code to the clipboard
uses xdotool to paste it back to the original window

# Mic only, 16 kHz mono, voice filtering, fast Opus
ffmpeg -hide_banner -v error \
  -f pulse -i default \
  -ac 1 -ar 16000 \
  -af "highpass=f=100,lowpass=f=6000" \
  -c:a libopus -b:a 16k -vbr on \
  -compression_level 2 -application voip -frame_duration 60 \
  -y "$AUDIO"

# Transcribe, extract the code fence and copy to clipboard
llm -m gemini-2.5-flash -a "$AUDIO" -s "$SYSTEM_TEXT" \
  | tee /dev/tty \
  | awk 'BEGIN{f=0} /```/{f=!f; next} f{buf=buf$0"\n"} END{print buf}' \
  | xclip -selection clipboard

# Bring last window to foreground and paste from clipboard
if [ -n "${ACTIVE_WIN:-}" ]; then
  xdotool windowactivate --sync "$ACTIVE_WIN"
  sleep 0.08
  xdotool key --clearmodifiers ctrl+shift+v
fi

During the workshop, I said, “Which (judicial) bench has the longest pending cases,” and it generated a DuckDB query that, single-shot, ran correctly on the Indian High Court judgements dataset.

But LLMs are slow and break the flow. Here’s how I keep the room engaged:

Dictate, don’t type. Speaking is faster and more engaging. That’s why I built this workflow.
Avoid Alt-Tab. Bring the LLM into your app. Window-switching and copy-paste break focus for you and the audience.
Always stream output. Narrate as it loads. | tee /dev/tty streams while piping.
Answer questions while you wait. Keep a Slido Q&A open and address the top ones while waiting for the LLM.

Things I Learned - 03 Nov 2024

Sun, 03 Nov 2024 00:00:00 +0000

This week, I learned:

Indian companies with 30+ employees MUST have 2.5%-15% of their employees as apprentices. Ref
Textnow and TextFree provides a free phone number (like a virtual SIM). (But TextFree has more ads.) Keep using to avoid deactivation. No guarantee of retaining the number.
- Some banks don’t accept TextNow for verification SMS. But voice call is OK.
- Tello, Red pocket are cheap MVNOs with $5/month voice plans.
- Metro by T-Mobile and Cricket are other MVNOs.
- MintMobile and US Mobile have $15/month and $8/month data plans.
The scientific discoveries that might have remained undiscovered for long if not for their discoverers Ref
- Newton’s discovery of the universal law of gravitation
- Einstein’s discovery of General Relativity
- McClintock’s discovery of Transposable Elements: genes that can turn physical characteristics on and off
- Mullis’ invention of the PCR that makes billions of DNA copies rapidly
VibeCheck can predict a model based on its vibes 80% of the time.
/llms.txt is a proposal to standardize /llms.txt files as a way to share LLM prompts.
- Jina AI Meta Prompt is an example
- Remotion system prompt is an example
- https://docs.fastht.ml/llms-ctx.txt
- https://docs.fastht.ml/llms-ctx-full.txt
structuredClone deep clones objects in JS
F5-TTS clones voices with just 15-second samples.
Rust has crazy low memory usage too. Spawning thousands of child processes is common and OK these days. Ref
SetInterval is a good idea in cyborg scraping. Ref
GH CLI is quite good for deployment too, like Wrangler CLI. Enabling pages, setting secrets, etc.
Restic is a CLI backup tool. Just like git. Works well with rclone.
NotebookLlama is an open source podcast generator like NotebookLM
Pragmatic Podcast (I forgot which one)
- Automate changelogs for your codebases. Convert past commits into attractive release notes automatically
- AI is going to be the consumer of many tools and logs. Build converters for these
- Speed of validation such as linting, testing, etc. will allow LLMs to iterate faster and WILL become more important
Via Soumya Ranjan
- Vision embedding is useful in agile modeling
- Vision embedding models with SAM, Grounding Dino by meta, Alibaba does good stuff
- Vision embedding is more useful in batch than real time
- Embedding subtraction with vision embedding models like Dino
AI code editors are not good with large code bases today. Keep the refactoring exercises to below 1000 lines. Also evaluate the ease of setting it up locally
Deepseek Janus is a 1.3b model that can generate both text AND images (and also supports vision)
Cohere Multimodal Embed v3 is available on Azure.
Elevenlabs lets you create voices with a prompt. No need to even clone one!
Runway Act One creates expressive character performances

Clone any voice with a 15-second sample

Thu, 24 Oct 2024 01:36:21 +0000

It's surprisingly easy to clone a voice using F5-TTS: "A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching".

Here's a clip of me, saying:

I think Taylor Swift is the best singer. I've attended every one of her concerts and in fact, I've even proposed to her once. Don't tell anyone.

(Which is ironic since I didn't know who she was until this year and I still haven't seen or heard her.)

You'll notice that my voice is a bit monotic. That's because I trained it on a segment of my talk that's monotonic.

Here's the code. You can run this on Google Colab for free.

A few things to keep in mind when preparing the audio.

Keep the input to just under 15 seconds. That's the optimal length
For expressive output, use an input with a broad range of voice emotions
When using unusual words (e.g. LLM), including the word in your sample helps
Transcribe input.txt manually to get it right, though Whisper is fine to clone in bulk. (But then, who are you and what are you doing?)
Sometimes, each chunk of audio generated has a second of audio from the original interspersed. I don't know why. Maybe a second of silence at the end helps
Keep punctuation simple in the generated text. For example, avoid hyphens like "This is obvious - don't try it." Use "This is obvious, don't try it." instead.

This has a number of uses I can think of (er... ChatGPT can think of), but the ones I find most interesting are:

Author-narrated audio books. I'm sure this is coming soon, if it's not already there.
Personalized IVR. Why should my IVR speak in some other robot's voice? Let's use mine. (This has some prank potential.)
Annotated presentations. I'm too lazy to speak. Typing is easier. This lets me create, for example, slide decks with my voice, but with editing made super-easy. I just change the text and the audio changes.

Things I Learned - 03 Mar 2024

Sun, 03 Mar 2024 00:00:00 +0000

This week, I learned:

You can use slots to stream HTML out of order!
Shane Parrish. Short-term patience podcast
- have a frame of reference to relate EVERY experience to. That helps you evaluate (measure) and learn. That’s part of what Charlie Munger’s lattice of frameworks is about
- when there is a very high or very low interest scenario, low interest scenario then go ultra long term. Issued hundred years when the interest rate regime was very low
- short term optimal is rally long term optimal. So you need to learn to take a loss and look like an idiot to play the long-term game
- grit is a behavior that enables long-term thinking. Short term success gives you the luxury to think about long term
- #IMP power is about optionality. It’s about being in a position where you have the options that can affect the positive change rather than circumstances controlling you. Read Robert greene’s book on the 48 laws of Power
- low leverage enables that
- begin with the end in mind. Always
- how do you think about risk? Well, things do happen. It’s as simple as that
- autonomy and decentralization helps derisk
- do more and more of what works. That’s a powerful way of compounding
- long-term investments are better than frequent trading because you get to reinvest the tax you otherwise would have paid. So unless the alternative is super compelling, stay invested
- if you need to be the person who DOES the thing, you delegate less, leverage list, compound less, because you have to DO. BE A PERSON WHO SETS THE FIELD INSTEAD. The coach, the chess master, the director, patient strategist who Waits for the good hit
Being in Control motivates #Lesson. my cycle tires were flat. I thought it was someone pulling out the air and felt very demotivated. But once I carried my cycle pump, I felt so much more in control and power and felt a whole lot better
SourceGraph is the default platform for private code completion & search
MetaVoice 1B offers voice cloning on American & British accents with 30s training
Qwen 1.5 72B appears to outperform Mistral Medium, making it one of the top non-proprietary models
Llava 1.6 is a substantial improvement over Llava 1.5 and slightly better than CogVLM, Qwen-VL
AI scams are growing. Deepfakes scammed $34m. But voice fake for kidnapping is scarier.
Buildspace’s demo is a great demo of how voice and actions can be used effectively.
demucs does an EXCELLENT job of splitting songs into drums, bass, vocals and others