rakeaudio-asr Available now
Conversations are documents too.
Two fine-tuned models in one pipeline. A recording goes in, a transcript with speaker labels and word timings comes out, and it enters the same pipeline as a parsed PDF.
- Transcription
- 270x
- Realtime. A 10-minute recording comes back in 2.2 seconds.
- With speakers separated
- 150x
- Same file, 3.9 seconds. An add-on at every volume, not a higher tier.
- Word error rate
- ~5%
- Measured internally on the Artificial Analysis English methodology.
- Per 1,000 minutes
- $5.00
- One flat rate at every volume. No minimums.
01The business case
rakeaudio-asr
A transcript nobody reads is not a record.
The value of a recording is in what happens next, and what happens next has a window. Miss it and the transcript becomes a chore for the end of a shift.
02The model
rakeaudio-asr
Audio does not hold still.
An invoice repeats its layout and a contract has a structure you can rely on. A conversation goes wherever it goes, and the pipeline has to hold up anyway.
- How many people are talkingA plain transcription API One block of text. Work out who said what yourself, if the model offers speakers at all.rakeaudio-asr Speaker labels and word timings, on a recording where people join and leave partway through.
- Which language it is inA plain transcription API A configuration field, pushed back onto you, on a file nobody has listened to yet.rakeaudio-asr Language detection is included. We are in Montreal, where a sentence can start in one language and end in another.
- How long you waitA plain transcription API Fourteen to twenty-one seconds on a ten-minute file, which is a batch job with extra steps.rakeaudio-asr 2.2 seconds, or 3.9 with the speakers separated. Five to ten times faster than the OpenAI models we measured.
- What you get at the endA plain transcription API A text file you have to go and find later, in a service that does nothing else.rakeaudio-asr A document in the same API as parse, extract, compose, redact and sign, ready for the next step.
03Specification
rakeaudio-asr
Everything it is, on one page.
One pipeline, one endpoint, one rate. Audio is heavier than paper, so the run is asynchronous and says so rather than holding a connection open.
Model
- Name
- rakeaudio-asr
- Category
- Speech recognition pipeline Two models we fine-tuned and combined, behind the Parse capability.
- Task
- Transcription and speaker separation Word timings and language detection are included, not add-ons.
- Languages
- Detected, not configured Including a recording that changes language partway through a sentence.
Speed
- Transcription
- 270x realtime 2.2 seconds on a 586-second two-speaker file.
- With diarization
- 150x realtime 3.9 seconds on the same file.
- Accuracy
- About 5% word error rate An internal measurement on the Artificial Analysis English methodology, not a leaderboard submission.
Interface
- Endpoint
- POST /v1/parse The same call that reads a PDF. A recording is another kind of file.
- Modes
- transcribe, transcribe_diarize Set on the file. Use the first when you do not need speaker labels.
- Execution
- Asynchronous Fire the request, then poll the run until it leaves processing.
Price
- Transcription
- $5.00 per 1,000 minutes The same rate at every volume, priced per minute of audio rather than per request.
- Speaker separation
- Add-on On top of that line, at every volume. The rate is on the pricing page.
04Benchmarks
rakeaudio-asr
Fast enough to stay in the conversation.
There are workflows where transcription quality is the whole job and nobody is waiting. Then there are the ones where the transcript is the next step in a conversation that just happened. This model is built for the second kind.
- 270 x Realtime, transcription A 10-minute recording in 2.2 seconds
- 150 x Realtime, diarized The same file in 3.9 seconds, speakers separated
- ~5% Word error rate Within a point of Deepgram, measured internally
- $5.00 Per 1,000 minutes One flat rate at every volume, no minimums
Speed on a 10-minute file, as a multiple of realtime
- Deepgram Nova-3 300
- rakeaudio-asr ours, transcription 270
- rakeaudio-asr, diarized ours, speakers separated 150
- Gemini 3.5 Transcribe 87
- OpenAI gpt-transcribe 42
- OpenAI gpt-4o-mini-transcribe 41
- OpenAI gpt-4o-transcribe 37
- OpenAI whisper-1 29
Batch and pre-recorded audio. Our rows come from an internal run at high utilisation with no cold starts, on a 586-second two-speaker file. Hosted rows come from vendor pricing pages and the Artificial Analysis speech-to-text leaderboard, measured on a 10-minute clip. Gemini 3.5 Transcribe is the most accurate pipeline in this set at 2.6% word error rate, and Deepgram is faster than we are. We would rather you run your own recordings than take any of these numbers on faith.
05The hard part
rakeaudio-asr
Which version of a fact is the true one?
In a document, the final agreed form of something is usually right there on the page. In a conversation it is not. Ideas get raised, reshuffled and quietly dropped, and the version that counts is either the most recent statement or the one nobody objected to.
No transcript settles that on its own. What a transcript can do is give the next model a clean, attributed, timed record to reason over, which is why the speakers and the timings are part of the output rather than an upgrade.
- Speakers join and leave inside the same recording, and the count is never known in advance.
- A sentence can start in one language and end in another, so the language cannot be a configuration field.
- Word timings put every claim back on the recording, so a reviewer can go and listen.
06Pricing
rakeaudio-asr
One rate, whatever the volume.
Priced per minute of audio rather than per request, so a thirty-second voicemail and a two-hour board meeting cost the same per minute. No tiers to negotiate into and no minimums to clear.
$ 5.00
per 1,000 minutes transcribed. That is $300 for a thousand hours of meetings, with word timings and language detection included.
- A thousand hours of meetings
- $300
- Speaker separation
- add-on, any volume
- Word timings and language detection
- included
- Minimums
- none
07Integration
rakeaudio-asr
A recording is another kind of file.
Transcription is the same call that reads a PDF. Send a file, get back a transcript with speaker labels and timings, then pass it to any other capability in the same pipeline. Audio runs asynchronously, so fire the request and poll the run until it leaves processing.
curl -X POST https://api.cloudraker.com/v1/parse \
-H "Authorization: Bearer $CLOUDRAKER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"file": {"url": "https://example.com/consult.m4a", "processing": "transcribe_diarize"}}'
# Or with the CLI. Use "processing": "transcribe" when you do not need speaker labels.
paperwork parse --wait 0 \
--json '{"file": {"url": "https://example.com/consult.m4a", "processing": "transcribe_diarize"}}'08Questions
rakeaudio-asr
Important questions, fast answers.
Is the ~5% word error rate a benchmark submission?
No, and we say so every time we quote it. It is an internal measurement using the Artificial Analysis English methodology, not a leaderboard run. Gemini 3.5 Transcribe is more accurate at 2.6%. Run your own recordings before you commit.
Is anything faster?
Deepgram Nova-3, at roughly 300x realtime against our 270x. We are faster than everything else we measured, by five to ten times against the OpenAI models.
Do I have to tell it what language the recording is in?
No. Language detection is included, including recordings that change language partway through a sentence. Asking the customer to configure a language for a file nobody has listened to yet is not a workable ask.
Why is speaker separation an add-on rather than included?
Because it costs real time and real compute, and plenty of workflows never need it. Making it an add-on at every volume keeps the base rate flat instead of pricing diarization into every minute you transcribe.
Can I use the transcript with the rest of the API?
That is the point of it. A transcript enters the same pipeline as a parsed PDF, so extract, compose, redact and sign all work on it without a separate integration.
rakeaudio-asr is one of three.
Same infrastructure under each one. Pick the model that fits the input you actually have.
Send us the recordings that break it.
Two speakers talking over each other, a room with bad acoustics, a call that switches language mid-sentence. That is how the next version gets trained.