Mimicking the Master: How I Trained an AI Bot to Speak Like Harsha Bhogle.

Harsha Bhogle

Cricket is more than just a sport in countries like India; it's a shared emotion, a cultural thread that binds millions. And for many fans, the voice that has narrated this journey for decades is none other than Harsha Bhogle. Known as “the voice of Indian cricket,” Harsha is celebrated not just for his deep understanding of the game but for his articulate, insightful, and poetic commentary that brings every match to life. His ability to blend analysis with storytelling has made him a beloved figure across generations. In this project, I set out to create an AI bot that could emulate Harsha Bhogle’s unique voice, a blend of cricket expertise, elegance in language, and journalistic balance, and in doing so, learned a great deal about the interaction of technology, linguistics, and personality modeling.

The core of this project relied on designing a strong persona through prompt engineering, tone conditioning, and fine-tuning example dialogues.

The first step was defining a system prompt that would act as the personality blueprint for the bot. The system prompt used in OpenAI’s GPT-based interface looked something like this:

system prompt

The prompt was critical in setting the bot’s behavior. By conditioning the system like this, every response was generated within the bounds of Harsha Bhogle’s persona.

Next I compiled example interactions based on Bhogle’s real-life commentaries, tweets, and interviews. For example:

Example

The example was used during testing to calibrate tone, vocabulary, and the style of analogies Bhogle often employs. To ensure consistency, I used few-shot learning techniques, feeding several such examples to the model during each prompt session. Technically, it worked within a chatbot interface powered by OpenAI, integrating contextual memory to handle follow-up questions.

One challenge was to avoid generic or robotic replies, which I solved by pre-filtering certain keywords, encouraging dynamic sentence structure, and tuning the bot to prefer narrative-style answers over bullet points.

In conclusion, creating this bot taught me how to blend prompt engineering, tone design, and persona training to create a believable AI version of a real person. It wasn't just about mimicking speech but was about recreating a character.

In the year 2022, I remember OpenAI posted a video of Joe Rogan interviewing Steve Jobs, and I was amazed; it was nothing short of magic for me at that time, but now I understand the science and technical aspects of it. This is just a beginning. Imagine if you could create a persona of your loved ones and add voice capabilities to it with a face movement and store it in a frame of a kind or a web page that you could refer back to if you ever needed advice or just felt like talking to them. Imagine the possibilities.