Snipd

Snipd is a Zurich-based company building an AI-powered podcast app for learning. Listeners can capture a moment through their headphones, saving its audio, transcript, and an AI summary, then revisit episodes through searchable transcripts and chat. Its March 2026 AI DJ release extended the app into guided listening: it plays selected passages of original audio and adds spoken context between them. These features grew from an earlier podcast-search prototype and a product initially organized around shared clips.[1][4][9]

Snipd's origins precede its note-taking and guided-listening releases. It launched after taking first place at HackZurich 2020. In a March 2025 interview, Kevin Ben-Smith recalled building natural-language search within podcast episodes with a friend who later became a co-founder. Their demonstration searched for “smoking weed” in Joe Rogan's long interview with Elon Musk and located the relevant audio moment; they used video for the stage presentation. Other participants subsequently asked to use the prototype and described related problems, helping convince the pair to leave their jobs and start Snipd.[4][1]

Ben-Smith brought experience applying machine learning to banking. He recalled studying mathematics and economics with a quantitative-finance specialization at ETH Zurich, discovering machine learning through an ETH course recommended by the same friend, and leaving a bank job for an early-stage Zurich startup. Over five years there, he built up its AI team, working on sales recommendations and transaction-text processing. That work supplied practical AI experience before the podcast-search experiment.[1]

Snipd was founded in 2021 by Kevin Smith, Ferdinand Langnickel, and Mikel Corcuera. Patricia Allen's March 3, 2022 EU-Startups account reported that it had raised €631,000 in an oversubscribed pre-seed round, led by Wingman Ventures with participation from Acequia Capital and Contovista co-founder Gian Reto à Porta. The stated plans were product development, team expansion, and establishing product-market fit. The financing account names Kevin Smith as co-founder and CEO; the 2025 interview introduces its guest as Kevin Ben-Smith.[4][1]

The early product combined personal capture with discovery through short podcast highlights called snips. Ben-Smith recalled that the team initially envisioned a social, TikTok-like feed, expecting listeners to consume clips readily but create them reluctantly. It therefore concentrated on making capture easy: a headphone triple-tap would save and summarize the moment just heard. In his account, users instead “were creating Snips like crazy.” That observation encouraged the team to emphasize learning and knowledge capture. The workflow addressed listening while driving, exercising, or walking a dog—situations where writing a note or recording a timestamp is awkward.[4][1]

The October 11, 2024 redesign made saved insights more editable. Separate title and summary fields became one unrestricted text field, and an AI rewrite feature let users request a different focus, more context, or another format. Newly processed transcripts gained speaker names. The release also replaced the “For You” feed with episode discovery based on community snipping, offering trending, last-30-days, and all-time views. The replacement feed was not personalized at launch.[5]

Two November 2024 releases broadened capture and retrieval. Chat with Episodes, announced November 8, let users open an episode in their library and ask questions against its transcript, receiving answers with timestamps. Uses included finding a discussion, collecting quotes, and summarizing a topic spread across an episode. Auto-Snipping, announced November 26, optionally created notes with transcripts and summaries during listening. Manual headphone capture remained available, and the library could distinguish AI-created snips from personally selected ones.[6][7]

In March 2025, Ben-Smith described a four-person technical team, split evenly between backend and AI work and frontend development. Approximately 90% of the backend was Python, hosted on Google Cloud Platform; Flutter and Dart supplied a shared Android and iOS codebase, while the Apple Watch app was native. The AI pipeline had evolved from self-hosted open-source models and attempted fine-tuning to Whisper transcription and downstream processing mainly with OpenAI and Gemini models, using Perplexity for web search. Some transcription and speaker-diarization models remained self-hosted.[1]

Speaker identification combined diarization—grouping speech belonging to the same speaker—with transcript content, guest information, and LLM processing to assign names. The team modified clustering and used podcast-specific heuristics: someone speaking for only 30 seconds in an hour-long episode might be an advertisement voice rather than a host or guest. It also combined transcripts and LLMs with diarization to adjust speaker boundaries; Ben-Smith acknowledged that results remained imperfect. A separate synchronization algorithm addressed dynamically inserted advertisements, which can change an episode's audio file and invalidate transcript timestamps. Fuzzy matching at the audio level realigned playback with the transcript.[1]

The team paired model output with narrower engineering controls. Timestamp-linked chat answers used prompts to specify formatting and regex corrections to repair streamed responses. For the year-end 2024 recap, Ben-Smith recalled replacing occasional dull quote selections with a two-step process: a cheaper model generated five candidates, and a stronger model chose one. Other LLM checks challenged book and speaker extraction. Hosted APIs supported rapid iteration and variable, real-time user requests, while episode preprocessing ran through a predictable, continuous queue. Experiments with Gemini 2.0 Flash processing raw audio had not replaced the existing transcription and diarization pipeline because of cost.[1]

On December 23, 2025, Snipd announced a Libro.fm integration extending headphone capture, transcripts, summaries, and chat to audiobooks. Libro.fm supplied purchased DRM-free files: users logged into their account within Snipd, viewed their purchases, and uploaded a selected book. Audiobook highlights could synchronize automatically to Readwise, Notion, and Obsidian, with manual exports to Logseq, Bear, and other Markdown-based apps. Snipd also described a connected workflow from a book recommendation in a podcast, through its Mentioned Books identification and Libro.fm purchase link, to listening and highlighting the audiobook.[8]

The March 26, 2026 AI DJ launch added a narrated route through an episode. It introduces the episode, seeks to and plays selected original passages, supplies short spoken transitions, and ends with a wrap-up. Listeners can exit DJ mode and return to normal playback. Launch coverage included English-language podcasts and uploaded English audio, including audiobooks, YouTube videos, and recordings. Snipd advertised roughly 25% of the original listening time, illustrating this with a 60-minute episode taking about 15 minutes; that figure is the company's product claim, rather than a measured learning or retention result.[9]