Practical guides on offline transcription — and a deep field guide to how speech recognition and on-device AI actually work.
Hands-on, vendor-honest guides to getting clean transcripts from real-world audio.
The most common cause of a bad transcript isn't the software — it's the input file. Here is how to turn any recording into the clean 16 kHz mono WAV that speech-to-text engines want, with copy-paste ffmpeg commands — and an honest note on when you can skip all of it.
Read article →Most "best transcription app" lists are written for phone voice memos. This one is for desktop work: long recordings, multiple speakers, batch jobs, and audio that cannot leave your machine. Here are the offline tools that hold up, what each is best at, and where each falls short.
Read article →Getting good audio in — and good transcripts out of imperfect, real-world recordings.
When a transcriber records your mic and the far side of a call at once, it quietly cancels acoustic echo — separating two mixed voices using the one thing that makes it possible: a reference copy of the far-end signal. The elegant idea behind clean two-sided recording.
Read article →A model trained only on clean studio speech falls apart on a real recording. The fix is to deliberately degrade the training data — noise, reverb, speed, SpecAugment — so the model expects imperfection. A quiet reason offline transcription survives real-world audio.
Read article →A field guide to ASR — one technique per post, from the classical Kaldi stack to Whisper and on-device models.
Speech-to-text got here by shedding complexity in stages — from room-sized statistical pipelines to a single model that runs offline on a laptop. A map of how ASR evolved, and the index to our deep-dive series on each technique.
Read article →Before CTC, training a recognizer meant telling it exactly which audio frame was which sound. CTC (2006) removed that constraint and became the engine of end-to-end ASR from 2014 to 2020 — a hinge in how speech recognition evolved toward today's on-device models.
Read article →CTC made recognition end-to-end but had no memory of what it just said. RNN-T (2012) fixed that and could stream — the architecture that put real-time, high-quality speech recognition on phones, offline.
Read article →A different idea from CTC and RNN-T: read the whole utterance first, then write the transcript while attending to the slice of audio that matters for each word. How Listen, Attend and Spell (2015) brought attention to speech.
Read article →In 2017 the Transformer threw out recurrence entirely — attention all the way down. It scaled where RNNs choked, and it is the architecture under Whisper and nearly every modern speech model. How it works, and why it changed ASR.
Read article →The Transformer was built for text; speech is also intensely local. The Conformer (2020) folds convolution back into the Transformer to give it a local ear — and became the acoustic workhorse behind a great deal of modern ASR.
Read article →For two decades before end-to-end models, speech recognition was a pipeline of specialized parts — features, a GMM-HMM acoustic model, a lexicon, a language model — and Kaldi was what ran it. The baseline the whole evolution measures against.
Read article →How do you search the astronomically large space of possible transcripts fast enough to be useful? The classical answer: compile the HMM, context, lexicon and language model into one weighted graph — HCLG — and walk it.
Read article →Two people saying the same word make very different audio. The iVector (2011) reduced a voice to a small fingerprint so classical ASR could adapt to the speaker — and became the backbone of speaker verification and diarization.
Read article →Every model so far needed labeled audio, which is expensive and scarce. wav2vec 2.0 (2020) borrowed BERT's masking trick to learn speech from oceans of unlabeled audio first, then fine-tune on a fraction of the labels — cutting the requirement ~100x.
Read article →A pipeline of hand-built boxes collapsed, era by era, into one model. In 2022 Whisper finished the job — and was small enough to run on the machine that recorded the audio. Where the whole arc lands, and why offline transcription is now good enough to trust.
Read article →The foundations beneath modern speech AI, and why capable models now run entirely on your own machine.
A survey we wrote in 2018 split AI into online (ship data to a server) and offline (run it on the device). Offline meant privacy — and, then, tiny models. In 2026 it no longer means tiny models, and that single change is the whole story.
Read article →AI, machine learning, and deep learning are nested rings, not synonyms. A short map: the classical algorithm toolbox, what a neural network actually is, and how neural nets lost to SVMs in the 1990s then won everything back around 2012.
Read article →Almost every model in modern AI — from a line through house prices to Whisper — is trained by the same humble loop: make a guess, measure how wrong it is, nudge it less wrong, repeat. That loop is gradient descent, explained clearly once.
Read article →