Turkish is a good Whisper stress test, agglutination makes word boundaries hard and smaller models fumble the suffixes. Large-v3 handles it far better than v2 did. One tip for non-English: set the language flag explicitly instead of auto-detect, which drifts on code-switching.

Parakeet trades accuracy for speed and size, so on hard audio it underperforms Whisper. Whisper small or medium on-device is the sweet spot: slower than Parakeet but far steadier on accents and noise. If the hardware has the RAM, medium is the better call despite the latency.

That volume works because podcast audio is single-speaker and clean, the easy case for ASR. The cost curve bends hard the moment you add crosstalk, accents, or domain jargon, which is why meeting and interview transcription still lags podcast transcription by a wide margin.

For heavy note-taking on iPhone, built-in dictation is underrated for live capture but it drifts past a minute or two. For long sessions, record and transcribe after: an app that batches the file handles hours cleanly where live dictation degrades. Depends if you need it real-time.

The local Whisper build winning here is usually about VAD and chunking, not the model. The API splits long audio on silence and can drop words across boundaries; running local lets you set chunk size and overlap. For botched audio, smaller chunks with overlap recover what the API skips.

For infrequent use, per-file pricing beats a subscription, exactly the Descript trap you hit. BrassTranscripts is pay-per-file, no monthly, with speaker labels, solid on interview audio. If cloud is fine, free tiers exist too. Match the tool to how often you actually record.

Diarization clusters on voice-print features, pitch, timbre, cadence, so a genuine shift in vocal quality can read as a new speaker. It's usually tunable: give it a known speaker count up front, or raise the merge threshold, and the over-splitting collapses. The default guesses generously.

Curious which diarization backend they're using under the hood since WhisperKit itself doesn't ship speaker separation natively. If it's pyannote bolted on, accuracy will hinge on manually setting expected speaker count rather than auto-detecting it, especially on calls with crosstalk.

Otter works but burns through your free minutes fast on a whole course. If you can save the lecture audio, running it through a local Whisper model gives unlimited passes and usually tracks timestamps better than live capture, which tends to lag on fast talkers.

Happens because the model's language layer overrides rare tokens with statistically common ones from training data. If your tool supports a prompt or custom vocabulary field, drop the Greek names in there before each run. Cuts that kind of substitution a lot.

For lecture recordings specifically, watch for hallucinated repeated phrases during long silent stretches, like when you're writing on a board or waiting for questions. Running a VAD pass to strip dead air before feeding audio to Whisper cuts that down a lot.

Silence and long pauses in interview audio are the real test for local Whisper builds. Without a VAD pass first, it'll hallucinate repeated phrases or invented text during quiet stretches. Matters a lot when IRB wants verbatim accuracy, not just speed.

Pyannote's clustering step often falls back to CPU on Apple Silicon since some torch ops aren't implemented for MPS yet, so diarization ends up slower than the 14B inference itself. Chunking audio into 20-30s segments before clustering cuts that bottleneck a lot.

Check whether Izwi lets you set keyword-bias phrases per request, not just bake them into a static config. That's the gap between real use and a demo feature. Also test Granite's diarization against pyannote on noisy multi-speaker audio before trusting it in production.

Most of that garbage comes from hallucinated text on silent or noisy stretches, not misheard words. Running a VAD filter to strip dead air before feeding audio in cuts hallucinations a lot. An initial_prompt with your domain vocab and names also helps a ton.

One gotcha that guide should flag: running WhisperX diarization alongside large-v3 on a single GPU often blows past 10GB VRAM since the pyannote model loads on top of Whisper. Loading them sequentially and freeing CUDA cache between steps fixes most OOM crashes.

Diarization accuracy drops fast with overlapping speech or similar voices, so WhisperX's forced-alignment step matters as much as the ASR model itself. Running pyannote 3.1 for VAD before segmenting cuts misattribution a lot on crosstalk-heavy interviews.

Descript's auto-transcribe falls apart on crosstalk and needs manual speaker fixes after every episode. Otter is fast but timestamps drift on anything over 40 minutes. Raw Whisper gives the cleanest text but has zero speaker labels, so you still need pyannote for diarization.

WhisperX is the open-source pick most people land on. Runs locally on your own GPU, gives word-level timestamps plus speaker diarization, and there's no subscription since you own the compute. Tradeoff: setup takes more fiddling than Descript and a decent GPU helps speed a lot.