Scribe v2 for new sermons
Fastest here, nothing invented in the silence, the most sermon words, and the only engine to hear “Berean.” About $0.26 for this service; a subscription's monthly hours cover a weekly upload.
SPEECH TO TEXT / JULY 2026
How many words
fall through the cracks?
Compare transcription errors, throughput, API cost, and the open-weight models that can listen locally on your Mac.
00:04.2
The frontier is not one model.
00:08.7
It is the edge between cost speed and accuracy.
00:13.1
1 substitution / 20 words · 5.0% WEREvery dropped word changes the meaning.
MEASURED ON A SERMON / OCTOBER 3, 2026
One sermon, four engines.
Leaderboards test short, clean clips. A church service is 70 minutes of silence, singing, announcements, a Scripture reading, and a sermon. This is Remember Lot's Wife (Luke 17:26–32) from Berean Community Church, run end to end on an Apple M1 Max, 64 GB. The recording opens with 12:50 of near-silence. With no human transcript, the preaching (26:45–65:21) is scored as disagreement with Scribe v2, the closest engine to the other two.
Engine / runtimeWall timeInvented in silenceSermon wordsvs Scribe“Welcome to ___”Cost / hour
Scribe v2ElevenLabs API · Fastest and most complete. Word timestamps; formats references as “Matthew 11:23-24”.34s122× realtime, incl. upload0 wordsnothing6,073ref.Berean$0.22
Parakeet TDT 0.6B v3parakeet-mlx, 300 s chunks · Silent in silence, but skipped four sermon passages, including the Matthew 11 reading.73s57× realtime0 wordsnothing5,7427.7%Burning$0 local
Whisper large-v3-turbomlx-whisper, no carried context · Accurate once speech starts; hallucinates a phrase in every silent window.126s33× realtime52 wordsThank you. ×265,9423.8%Pering$0 local
Whisper large-v3-turbowhisper.cpp, -mc 0 · Same accuracy as mlx-whisper, slowest of the four.171s24× realtime57 wordsThank you. ×25 · “visit www.fema.org”5,9613.8%Brewing$0 local
Parakeet drops whole passages
- 38:02 There's no fear of judgment. Well, if He loves me unconditionally… 21s
- 57:00 The whole reason why Christ came, was crucified on the cross… 9s
- 57:40 Matthew 11:23–24: “And you, Capernaum, will not be exalted to heaven…” 19s
- 60:56 It's because we covet things that God said is going to be burned in the fire… 27s
Whisper still talks over silence
Turning off carried context stops the loop from spreading, but each silent 30-second window still gets its own “Thank you.” Keep Whisper behind a repetition filter, or as a fallback only.
NEXT ON THE BENCH
Open models released this year that could close the gap. WER is the Open ASR Leaderboard's English average, a different test from the AA-WER below. Parakeet v3 scores 4.86 there.
- Cohere Transcribe 03-2026 2B · 4.67%
Best open model on the leaderboard's long-form track (9.73 vs Parakeet's 10.72). Timestamp support unconfirmed; community MLX ports only.
- Qwen3-ASR-1.7B 2B · 4.31%
Lowest open-weight short-form WER; mlx-audio port. LLM decoder: test the silence. Chunk long audio yourself.
- MOSS-Transcribe-Diarize 0.9B · 4.64%
90 minutes in one pass with timestamps, speakers, and hotwords. LLM decoder: test the silence.
- Granite-Speech-4.1-2B 2B · 4.62%
Keyword biasing for names like Berean, Zoar, Capernaum. Best on segments up to about nine minutes.
2.2% WER · ElevenLabs
2.4% WER · 265.9× realtime
904.9× realtime · open weights
$0.37 / 1,000 minutes
THE ERROR MAP
Fewer mistakes, less friction.
Word error rate ↓ lower is better
2%
3%
4%
5%
6%
$0.25
$0.5
$1
$2
$4
$8
$16
API price per 1,000 audio minutes → log scale
HOSTED MODEL
MAI-Transcribe-1.5
Microsoft
- AA-WER
- 2.4%
- Speed
- 265.9×
- Price
- $6 / 1K min
- Frontier
- No
PRIVATE TRANSCRIPTION
Let your Mac do the listening.
Change memory and precision to plan a local setup. Availability varies by runtime, so treat these as weight-based estimates—not measured throughput on every Apple chip.
Best ranked fitVoxtral Mini
3.8% provider AA-WER
Model / useAA-WERMemoryLanguagesTimestampsLicenseFit
Whisper TinyOpenAI · 39M · The smallest useful baseline; ideal for on-device commands.—not in snapshot1.4 GB99SegmentApache 2.0ComfortableWhisper SmallOpenAI · 244M · A dependable multilingual choice for older Apple silicon.—not in snapshot1.7 GB99SegmentApache 2.0ComfortableParakeet TDT 0.6B v3NVIDIA · 600M · The speed leader; official weights now include MLX support.4.5%provider run2.1 GB25Word + segmentCC BY 4.0ComfortableWhisper Large v3 TurboOpenAI · 809M · Broad language coverage with a much smaller decoder than Large v3.4.6%provider run2.4 GB99Word + segmentMITComfortableWhisper Large v3OpenAI · 1.55B · The compatibility baseline with a mature local tooling ecosystem.4.1%provider run3.2 GB99Word + segmentApache 2.0ComfortableCanary Qwen 2.5BNVIDIA · 3B total · Strong English transcription plus optional transcript post-processing.4.3%provider run4.9 GBEnglishSegmentCC BY 4.0ComfortableVoxtral MiniMistral · 3B · Adds translation and audio understanding beyond plain transcription.3.8%provider run4.9 GB8SegmentApache 2.0ComfortableVoxtral SmallMistral · 24B · Best open-weight AA-WER here, but dramatically heavier than ASR-only models.2.8%provider run30 GB8SegmentApache 2.0Too largeBENCHMARK YOUR AUDIO
WER is the start, not the answer.
Use your domain
Meetings, calls, lectures, and medical dictation have different failure modes. Keep a human transcript set.
Normalize once
Standardize punctuation, casing, numbers, and hesitations before comparing word error rate.
Slice the errors
Break results down by language, accent, noise, speaker overlap, proper nouns, and audio duration.
Time the whole path
Include upload, endpointing, diarization, and post-processing—not just model inference.