how it works · the engine
What transcribes your episodes, named.
No mystery box. Every transcript on Podmenti is made by one specific, published speech model, running on hardware we own, and every AI Search answer comes from one named language model. Here is exactly what, and what that means for the text you get.
Speech to text: NVIDIA Parakeet TDT 0.6B v3
Transcription runs on Parakeet TDT 0.6B, version 3, an open speech-recognition model published by NVIDIA. It is a 600-million-parameter model built on the FastConformer architecture with a token-and-duration transducer decoder, which is the part that makes it both fast and precise about when each word was said. Parakeet models sit at the top of the open ASR leaderboard: on the standard English benchmarks they make roughly one error per sixteen words, on par with the best commercial engines, while running many times faster than real time.
- Word-level timing is native. The model emits a start time for every token as it decodes, so word timestamps are not estimated afterwards, they come out of the recognition itself.
- Punctuation and capitalisation are in the model. Sentences end, names are capitalised, questions get question marks, without a second pass.
- Twenty-five languages. Version 3 is multilingual, covering English and the major European languages. Our library is almost entirely English today, and English is where the accuracy figures above apply.
- Speech detection first. Audio is chopped at pauses by Silero VAD before recognition, so silence, music beds and long ad breaks do not turn into hallucinated words.
Where it runs
On our own machine, not a third-party transcription API: an 8-core server in Hetzner's Falkenstein data centre, running the model in quantised int8 form through ONNX Runtime. Two workers together transcribe about four to five minutes of audio per minute of wall time, which is how a typical hour-long episode is ready within the hour after you request it. Audio is fetched straight from the show's host, transcribed, and discarded.
The fast lane runs the very same pipeline, model files, voice detection and word timing included, packaged as our own container on a rented Nvidia L40S through Replicate: eggvelop/podmenti-parakeet, public, so anyone can inspect it. The GPU recognises the detected speech segments in large batches, which is why a three-hour episode takes only a little longer than a ten-minute one. The audio leaves our machine for the seconds it takes to transcribe it, and nothing is kept there. We measure every fast-lane episode from request to transcript and show the median on the plans page. It is priced at double because the GPU time costs us more.
Keeping it true over time
The same model does the re-alignment work when a host swaps ads and the audio shifts: we re-recognise short windows around the affected sections and move the timestamps to where those words now sit. How that works.
AI Search: Google Gemini Flash-Lite
AI Search sends the whole transcript to Gemini Flash-Lite, Google's fastest production model, with your instruction and a strict rule to return only verbatim quotes. It reads the entire episode in one pass rather than retrieving fragments, which is why "not found" means it looked everywhere. We use the current Flash-Lite release through Google's API and re-anchor every returned quote to the transcript ourselves, so the timestamps on an answer are ours, not the model's guess.
What it is not
- Not human-checked. These are machine transcripts. On clean speech they are excellent; on heavy crosstalk, thick accents or bad microphones they will have errors, and proper nouns are the usual casualty.
- No speaker labels yet. The text is not split by who is talking. It is on the list.
- Not edited. Filler words, false starts and repetitions stay in, because they are what was said.