Join Design Trial
PENTECOST SPARK

Pentecost Spark: when real-time multilingual transcription accuracy matters.

Pentecost Spark is the engine inside Pentecost. It sends each language to the speech and translation engines that tested best, cleans every line against your sermon manuscript, glossary and Bible, scores its own confidence, and retries what it doubts.

What happens to every line

Most live caption tools we've looked at run one engine end to end: speech to text, then machine translation. Spark does a lot more between the microphone and the screen.

How Pentecost Spark handles every line Your church's settings feed a route table that picks the speech and translation engine for each language from test results, checked against live health scores. Speech to text is primed with your glossary and manuscript names. Guards drop wrong-script output and echoed prompts. A domain pack adds glossary terms, Bible names and the cited verse in each language's own translation. The LLM cleaner fixes mishearings against your manuscript and translates every language in one pass. Confidence routing retries low-scoring lines on a stronger engine and keeps the words when confidence is very low. A backup engine is raced if the first one stalls. Captions go out with canonical terms enforced and stamped with the Spark version, and delay and dropouts feed back into the health scores. Where Spark chooses the engine No answer in time First valid answer Delay and dropouts feed the health scores Your church's settings Languages, tradition preset, glossary, sermon manuscript and Bible translation Route table Picks the speech and translation engine for each language from test results Health scores Live delay and errors for every engine and language Speech to text Engine chosen by the route, primed with names from your glossary and manuscript Guards Drop wrong-script output, echoed prompts and stock phrases before they show Domain pack Glossary terms, Bible names and the cited verse in each language's own translation LLM cleaner Fixes mishearings against your manuscript, translates every language in one pass, using the last few lines Confidence routing Scores every line. Low score: retry on a stronger engine with more context. Very low: keep the words, don't guess Caption out Your canonical terms enforced, stamped with the Spark version for the scorecard Backup engine Raced when the first engine stalls, so the line still gets a caption
Orange boxes are where Spark chooses the engine for each language and each line.

What makes Spark different

Six things a single engine doesn't do for you.

An AI cleaner on every line Live

Every line is cleaned and translated into every caption language in one pass, using your sermon manuscript, the last few lines and your church's tradition. When a line nearly matches the manuscript, it takes the manuscript's wording.

Heard: “in the face of such a trust”
Caption: “in the face of such atrocities”

Your glossary and your Bible Live

Your glossary primes the speech engine and is enforced on every caption, with known wrong forms replaced. When a verse is quoted, the cleaner gets the passage from your Bible translation in each language, so the caption matches the printed text.

October test: Bible names in the right form went from 91% to 98% on the lowest-cost engine. Sarai stays 撒萊, not 撒拉.

Confidence routing Live

Every line gets a confidence score. A doubtful line is retried on a stronger engine with more context. A line Spark can't trust is shown as heard, rather than “improved” into something wrong.

The best engine for each language Live

No single AI is best at every language. Spark's route table sends each language to the engine that tested best, watches live delay and errors, and races a backup engine when the first one stalls.

October test: Spark's route dropped 0 of 282 lines. Single engines dropped up to 35% in their worst language.

Domain packs Church: live

Glossary, speech hints, tradition presets, reference texts and checks, bundled for one field. Church comes first, with six tradition presets: none, Protestant, Catholic, Orthodox, Pentecostal and Progressive. Legal and conference packs can follow.

A version on every caption Live

Every caption is stamped with the Spark version and the route that made it, so any result on the Spark Bench scorecard can be traced to the exact setup that produced it.

The models Spark uses

The latest model from each company, each in the job it tested best at. When a newer model tests better, Spark moves to it in a new version.

CompanyModelWhat Spark uses it for
OpenAIGPT TranscribeSpeech to text
GPT-6 LunaCleans and translates most languages
GPT-6 SolBackup engine, and retries on doubtful lines
MicrosoftAzure AI TranslatorFills in any language the cleaner leaves out
AlibabaQwen 3.8 FlashCleans and translates Cantonese

Checked October 2026, Spark version spark-2026.10.1.

Early internal test

Why one engine isn't enough

In October 2026 we ran the cleaner stage on 282 lines in eight languages: Mandarin, Cantonese, Japanese, Korean, Thai, Tagalog, Indonesian and English. A line counts as dropped when no caption came back within 6 seconds.

Setup Lines dropped Worst language Cost per 1,000 lines
GPT-6 Luna alone 12 of 282 Cantonese: 35% dropped $0.49
Qwen 3.8 Flash alone 12 of 282 English: 20% dropped $0.16
DeepSeek V4.1 Flash alone 11 of 282 Thai: 18% dropped $0.19
GPT-6 Sol alone 1 of 282 English: 3% dropped $2.78
Pentecost Spark route 0 of 282 None $0.51

Each low-cost engine failed badly in at least one language. The one that didn't costs more than five times as much. Spark's route matched the reliable engine at the low-cost price.

One run, text input, costs at October 2026 list prices. On line-by-line translation quality the engines scored close to each other in a blind review. The full Spark Bench, with recorded church audio and blind scoring, is in progress.

How we'll prove it: Spark Bench

A single engine rarely fails on average. It fails in one language, or on names, and an average hides that. So Spark Bench reports every result per language, and leads with the worst one.

  • Same church recordings, with their manuscripts and glossaries, through every setup.
  • Against single engines as sold (Azure, Google, OpenAI), each with its own custom-vocabulary feature loaded with the same glossary.
  • Against Spark with one feature switched off, so you can see what the cleaner, the glossary and confidence routing each add.
  • Measured on dropped lines, name and term accuracy, translation errors per 100 lines (blind review against reference translations), speech accuracy, caption delay and cost per service hour.
  • Published in full: the method, the clips we have rights to share, the raw outputs and our scripts. Re-run every quarter.

Read the benchmark method →

Questions about Spark

Spark is the engine inside Pentecost that turns speech into captions. It picks the speech and translation engines for each language, races a backup engine when one stalls, cleans every line against your manuscript, glossary and Bible, scores its confidence and retries what it doubts.
The latest models from OpenAI (GPT Transcribe, GPT-6 Luna, GPT-6 Sol), Microsoft (Azure AI Translator) and Alibaba (Qwen 3.8 Flash). Spark chooses per language, from test results. The table above shows what each one does.
In the standard cloud setup, audio goes to the AI providers on each Spark route, and some of them process it outside Singapore. If audio must stay on site, a fully offline Pentecost Box is coming. A Singapore-only Spark profile is planned.
A fixed set of routes, packs and settings. Every caption records the version that made it, so a result on the scorecard can be traced to the exact setup that produced it.

Test Spark with your own recording.

Send us a recording of a service or talk, with its notes and glossary. We'll send back captions from Spark and from a single engine, side by side.

Send a recording