Longer Audios

[guiaugustinegri]

What is the relationship between the size of the content sent to podcastify and the "word_count" setting to get longer audios? I tested several settings but the maximum I could get was 4 minutes.

[souzatharsis]

LLMs are known for their limited output number of tokes size.
Long form podcast generation has become an issue of the highest priority since multiple requests have been made.

#166

I'll work on it next; No clue what the solution will be since there isn't any known best practice for it but it will be fun to develop a robust solution for it.

stay tuned

[guiaugustinegri]

Thanks for the answer. Today, the LLM that generates the podcast script is Gemini, right? What I do to get larger responses in LLMS is to use the OpenAI API, I make the request using the gpt4-o model and create a thread, which is a native option. This way, in addition to having the initial context, it can remember previous requests for each request, working similarly to the chatgpt app. So I ask it to send me the responses in separate blocks, so I can receive more tokens. I believe this could be an option to be tested.

[souzatharsis]

As of a couple of days ago you can choose 100+ llm models to generate the transcript including gpt4-o. however, in my experience, llms are trained to give the shortest answer possible i'm currently working on a new multi-prompt transcript generation approach leveraging a content chunking with contextual linking approach In that way, we generate detailed transcripts per input chunck and then stitch responses together while maintaining cohesion. Easier said than done, but I am almost there. Would love your feedback once done. Best Regards,

…

-- Thársis souzatharsis.com <http://www.souzatharsis.com/> <http://linkedin.com/in/tharsissouza>

On Sun, Nov 10, 2024 at 1:05 PM Guilherme Negri ***@***.***> wrote: Thanks for the answer. Today, the LLM that generates the podcast script is Gemini, right? What I do to get larger responses in LLMS is to use the OpenAI API, I make the request using the gpt4-o model and create a thread, which is a native option. This way, in addition to having the initial context, it can remember previous requests for each request, working similarly to the chatgpt app. So I ask it to send me the responses in separate blocks, so I can receive more tokens. I believe this could be an option to be tested. — Reply to this email directly, view it on GitHub <#168 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADTMY3LQY7CJUZ2A67TUVS3Z7575LAVCNFSM6AAAAABROLFFTOVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTCMRQGUYDANQ> . You are receiving this because you commented.Message ID: ***@***.***>

[guiaugustinegri]

Excellent! It's going to be awesome =)

[souzatharsis]

I've now managed to implement long-form podcast generation. Please see sample below.
Abstracting away the TTS model used (OpenAI), would love feedback on the long-form transcript generated!

Sample 1: open-notebook.ai and its sub-pages
TTS Backend: OpenAI

• Audio: https://drive.google.com/file/d/1isp2gnfjnf-tyhgi_FREIbf02EafYjK4/view?usp=drive_link
• Transcript: https://drive.google.com/file/d/1XMbD7bZtwqN5Hs-2ysmxeeNO2s6xMXbW/view?usp=sharing

Sample 2: Some tech docs on surfsense.net shared by the author here
TTS Backend: OpenAI

• Audio: https://drive.google.com/file/d/1MyLF6RcLg3o7lp1bZA9ZStO0oocdMIhB/view?usp=drive_link
• Transcript: https://drive.google.com/file/d/17nhnPpf_r3QzQ4Kq-8pgwQ_P6hYm_LfS/view?usp=drive_link

Sample 3: Transformer pdf paper
TTS Backend: OpenAI

• Audio: https://drive.google.com/file/d/1iYgMjqw7ct8yVCWS4oACQWEZJmsocYpT/view?usp=drive_link
• Transcript: https://drive.google.com/file/d/1LqCL26e_BtwuLzCE-eFzVGiisWCxknrz/view?usp=drive_link

[ralyodio]

Is this merged?

[ralyodio]

How do I specify length of audio?

[KlaasYntema]

I listened to all three of them and besides the feedback MalikMAlna provided, they sound pretty good!

[souzatharsis]

Thanks for the feedback! @KlaasYntema @MODSetter @MalikMAlna !

@ralyodio , it's not merged yet. I'll merge it once I collect enough feedback.

[toilv]

I still only get 7 minutes for file. geminimulti with --longform . Any suggesttion? I need file length at least 20 mins.

[souzatharsis]

Sample 4: Gemini Prompt Engineering Guide
TTS Backend: ElevenLabs

Transcript: https://drive.google.com/file/d/1ed2Xl9zPLyXPQNqJEcxGedgp9cMyN97T/view?usp=drive_link
Audio: https://drive.google.com/file/d/1jdnZC_3R0A9TNt9IPYZORz8_oT_o3YaX/view?usp=drive_link

[souzatharsis]

This has been implemented and released - I'd love feedback!

See https://github.com/souzatharsis/podcastfy/blob/main/usage/how-to.md#how-to-generate-longform-podcasts

[0.3.6] - 2024-11-13

Added

• Add longform podcast generation support
  • Users can now generate longer podcasts (20-30+ minutes) using the --longform flag in CLI or longform=True in Python API
  • Implements "Content Chunking with Contextual Linking" technique for coherent long-form content
  • Configurable via max_num_chunks and min_chunk_size parameters in conversation config
  • word_count parameter removed from conversation config as it's no longer used

[toilv]

I still only get 7 minutes for file. geminimulti with --longform . Any suggesttion? I need file length at least 20 mins.

[ralyodio]

I want at least an hour.