Accept the Kroko Community Model License 2.0 to access this model
Access requires recorded acceptance of the license. Commercial use is limited to licensees with less than US$1,000,000 in Group Revenue unless a Separate Agreement applies.
Submitting this form is the Official Acceptance Process under section 3 of the Kroko Community Model License 2.0 for the Covered Release Kroko-ASR-XL-EN-Offline. Your acceptance is recorded with the licensor together with the details below, your Hugging Face username and email address, and the date and time of submission. The licensor handles this information as described in its Privacy Notice. Acceptance is not consent to unrelated marketing.
Accept on your own behalf, or on behalf of one legal entity that you are authorized to bind. Each legal entity must register separately; an acceptance for one entity does not license its affiliates.
Kroko Community Model License 2.0
Copyright (c) 2026 Banafo EOOD
Licensor: Banafo EOOD, a company incorporated in Bulgaria
Registration number (UIC/EIK): 201141867
Registered office: 120 Tintyava Str., entrance B, 1st floor, Izgrev District,
1113 Sofia, Bulgaria
Licensing and legal notices: licensing@banafo.com
1. Definitions and scope
1.1 "Licensor" means the company identified above. "You" means the individual
accepting this License on their own behalf, or the legal entity expressly
identified when an authorized representative accepts it on that entity's behalf.
An acceptance for one entity does not automatically license its affiliates.
1.2 "Official Repository" means the model repository operated by the Licensor
in which this License is published. "Covered Release" means the model release
published there for which You accept this License, as identified in Your
acceptance record. "Original Model" means the weights and other model artifacts
that the Licensor publishes in the Official Repository as the Covered Release.
Separately licensed software, training data and third-party materials retain
their own terms, even if supplied in the same repository or package. "Model"
means the Original Model and any Permitted Conversion of it.
1.3 "Model Materials" means the Original Model and copies, conversions and other
versions made directly or indirectly from its weights, including fine-tunes and
merged weights. For the contractual obligations accepted under this License,
this term also includes model-specific learned adapters, parameter deltas and
additional or replacement prediction heads trained using the Model's internal
representations or integrated into its prediction path. Including an item in
this definition does not authorize its creation, use or distribution, or transfer
ownership of another person's contribution to the Licensor. The permissions for
modification are limited by section 2.
1.4 "Permitted Conversion" means a change made solely to deploy or run the
Original Model for inference, consisting of:
(a) changing serialization, file format, sharding or packaging, or applying
lossless file compression;
(b) exporting or compiling it for a runtime, including ONNX or TensorRT;
(c) changing numerical precision or applying post-training quantization; or
(d) applying inference graph or kernel optimizations, including operator fusion,
constant folding and equivalent runtime substitutions.
Calibration solely to select quantization scales, clipping, rounding, codebooks,
numerical representations or runtime settings is permitted. These operations may
introduce numerical differences inherent in conversion or quantization. They
must not include training or fine-tuning the Model, training a parameter adapter
or an additional prediction head, averaging or blending weights from models or
checkpoints, or adding learned task-specific capabilities. Quantization-aware
training and structural pruning are not Permitted Conversions. Removing
redundant inference operations without changing learned model structure is
allowed under paragraph (d). Repackaging or renaming a training procedure as a
conversion does not make it permitted.
1.5 "Authorized Personnel" means Your employees and contractors with access
solely to carry out permitted work for You, subject to confidentiality, access
controls and restrictions against independent use or onward disclosure. This
includes a cloud or infrastructure provider holding copies solely to operate
Your authorized environment. Your customers, affiliates and independent
collaborators are not Authorized Personnel merely because they work with You.
1.6 "Separate Agreement" means a written agreement with the Licensor that
expressly grants additional rights or changes specified conditions of this
License. Paying for a third-party product or hosting service is not, by itself,
a Separate Agreement with the Licensor.
2. Permissions and limits on modification
2.1 After completing section 3, and subject to this License, the Licensor grants
You a worldwide, nonexclusive, royalty-free license under the rights it controls
to download, run, evaluate and copy the Original Model; make and use Permitted
Conversions; and distribute or sell copies of the Original Model and Permitted
Conversions under section 5. Commercial activities must also satisfy section 4.
You may adjust inference settings, including decoding, beam search, batching,
streaming and contextual biasing, without training or changing learned Model
parameters. Independent software wrappers, search or decoding algorithms, and
external language-model rescoring of the Model's ordinary predictions are
permitted. This does not authorize adding or replacing a learned prediction
head under section 2.2. Independent components retain their own licenses.
2.2 This License does not permit fine-tuning, continued pretraining, retraining,
training model-specific adapters or parameter deltas, training or attaching an
additional or replacement learned prediction head, model merging, or other
modification outside the definition of Permitted Conversion. This includes
training a new CTC, classification or other prediction head on the Model's
encoder representations, even if all supplied weights remain frozen and the
new parameters are stored or distributed separately. Such activities require
express permission in a Separate Agreement. A commercial license that only
authorizes production use does not also authorize training or modification.
2.3 You may make private deployment copies, caches and backups, and provide
access to Authorized Personnel, without registering every internal copy or each
such person as a separate licensee. You remain responsible for their compliance.
The commercial rules applicable to contractors are stated in section 4.5.
2.4 Viewing or sharing this License, public model documentation, benchmark
results or independently licensed code does not require model registration,
provided Model Materials are not included. Permitted evaluation includes
publishing results and reproducible evaluation code.
3. Acceptance, registration and access
3.1 Before receiving access to Model Materials, You must expressly accept this
License through the Official Acceptance Process, through an application under
section 3.7, or through another process approved in writing by the Licensor that
records acceptance with the Licensor before access is provided. "Official
Acceptance Process" means the access request form that the Licensor operates at
the Official Repository. An acceptance on a distributor's own website or model
repository is insufficient unless that process satisfies this section.
3.2 The acceptance record must identify the individual or legal entity licensed,
a contact name and email address, the relevant country, the Covered Release and
License version, and the date and time of affirmative acceptance. For an entity,
the representative must identify its legal name and confirm authority to bind it.
You must have legal capacity to enter this License, supply accurate information
and keep the licensing contact current.
3.3 A valid existing acceptance may be confirmed through the Official Acceptance
Process or an approved verification method instead of requiring repeated
acceptance for copies or Permitted Conversions of the same Covered Release under
the same License version. A different legal entity must register separately,
and each Covered Release requires its own acceptance. Registration or platform
approval does not establish commercial eligibility or grant additional rights
beyond this License or a Separate Agreement.
3.4 Distributors must obtain confirmation that the required acceptance has been
recorded before providing access. Confirmation may be an acceptance receipt, an
approved verification result or written confirmation from the Licensor. Do not
require recipients to disclose platform passwords or personal access tokens to
prove acceptance, and do not share credentials to bypass another recipient's
acceptance requirement. For application end users, the record kept under
section 3.7 is sufficient confirmation.
3.5 The Licensor's collection and handling of registration information is
explained in the privacy notice that the Licensor makes available with the
Official Acceptance Process. An approved distributor must give the required
privacy information and lawfully transmit the necessary acceptance record.
License acceptance is not consent to unrelated marketing. This License does not
require transmission of recordings, transcripts or inference telemetry.
3.6 If the Official Acceptance Process is unavailable, contact the licensing
email above for an alternative recorded acceptance. An outage does not authorize
new unrecorded distribution, but does not by itself terminate an existing valid
license or prohibit existing compliant offline use.
3.7 If You distribute the Model only as part of an application, whether bundled
with it or downloaded by it, Your end users may accept this License within that
application instead of through the Official Acceptance Process. Before the
application first makes the Model available to an end user, it must present
this License or a link to it and obtain the end user's express acceptance. You
must keep a record of the License version and the date of each acceptance. The
end user does not need to register with the Licensor. An acceptance under this
section licenses the end user to run the Model within that application only. An
end user who extracts, redistributes or separately deploys the Model must first
complete the Official Acceptance Process. Section 4 applies to each end user's
own Commercial Use.
4. Commercial use and the US$1 million limit
4.1 You may use the Original Model and Permitted Conversions commercially under
this License while Your Group Revenue is less than US$1,000,000. At US$1,000,000
or more, You must obtain a Separate Agreement authorizing that commercial use
before starting or continuing it, unless the provider coverage in section 6
applies. Negotiations, registration or a breach cure period do not extend this
permission. There is no automatic grace period for crossing the threshold.
4.2 "Commercial Use" includes use in a business, internal business production,
paid consulting or model services, and providing or distributing a commercial
product or service, including a product offered free of charge. Charging for
redistribution, installation or a Permitted Conversion is Commercial Use.
4.3 "Group Revenue" means total worldwide revenue from all sources over the
immediately preceding 12 months, consolidated for You and entities that directly
or indirectly control You, are controlled by You or are under common control
with You. "Control" means the power to direct management or policies through
ownership, voting rights, contract or otherwise. Consolidation for this
calculation does not extend a license to another entity.
Calculate revenue before deducting operating costs, using the accounting
standards normally applicable to the relevant entities on a consistent basis.
Exclude intra-group transactions, sales taxes collected for public authorities,
refunds and returns, and equity contributions or loan proceeds that are not
revenue under those standards. For currencies other than US dollars, use the
average reference exchange rate for the period published by the European Central
Bank, or by the relevant reporting currency's central bank if no ECB rate is
available. Keep sufficient records to support Your calculation.
4.4 Check eligibility before Commercial Use, at least monthly thereafter and
promptly when a material revenue or control change may affect eligibility.
The threshold applies continuously; these checks do not postpone section 4.1.
4.5 When employees or contractors use the Model Materials for an organization,
that organization must qualify under this section or section 6. A contractor
must also qualify for its own Commercial Use, including paid model operation,
conversion or consulting, unless a Separate Agreement covers it. A contractor
requiring its own license must complete section 3 for its own entity. Merely
providing general-purpose computing or storage, without independently operating
or exploiting the Model, does not by itself require a separate model license.
4.6 An organization of any size may conduct non-production evaluation under
this License, after acceptance and registration, if the evaluation does not
deliver operational work to the organization or its customers. Evaluation does
not authorize training or modification prohibited by section 2.2. Noncommercial
use is not subject to the revenue threshold but remains subject to the other
conditions of this License.
5. Redistribution and downstream acceptance
5.1 Redistribution of the Original Model and Permitted Conversions is allowed
when You comply with this section and, where applicable, section 4. This includes
distribution through model hubs, mirrors, applications, containers, installers,
private delivery or physical media. The recipient must complete section 3 before
receiving access to the Model Materials, even if delivery is free of charge.
5.2 Every distributed copy must include this License, the applicable copyright
and other required notices, the Original Model's name and Covered Release, and
a link to the Official Repository and Official Acceptance Process. Identify
Permitted Conversions as modified versions and describe their material changes.
Do not imply that the Licensor created, approved or supports Your conversion
unless it has agreed to that representation.
5.3 Keep Model Materials behind an access mechanism that verifies section 3.
Do not place them in an unrestricted public download, repository, package or
container, or provide a shared credential or reusable public link that allows
unregistered recipients to obtain them. Public documentation and code may link
to the appropriate gated download. A separately gated mirror is not sufficient
unless its acceptance and verification process complies with section 3. An
application that delivers the Model only after acceptance under section 3.7
satisfies this section, provided the Model files are not also offered as a
separate download.
5.4 The commercial, acceptance and redistribution conditions continue to apply
to Permitted Conversions. Recipients receive rights in the Licensor's materials
directly from the Licensor under this License, not through a sublicense from
You. Each recipient that redistributes must satisfy these same requirements.
Charging for a copy does not remove the recipient's obligations.
5.5 You may separately license Your own code and contributions to the extent
You own the necessary rights, but must not remove or replace this License for
the Licensor's materials or grant permission to bypass its conditions. These
requirements do not transfer ownership of Your contributions to the Licensor.
Versions requiring permission under section 2.2 may be distributed only where
a Separate Agreement expressly authorizes their creation and distribution.
5.6 Keep sufficient evidence of compliant delivery, including the acceptance
confirmation and identity of the release or conversion supplied, and provide
reasonably requested, relevant evidence to the Licensor, subject to applicable
privacy law. This does not grant access to customer recordings, transcripts,
unrelated business records or systems. Retain personal information only as
lawfully necessary for these purposes.
6. Licensed providers and hosted services
6.1 A provider with a Separate Agreement covering a product or service may
extend the coverage expressly granted by that agreement to its end users,
regardless of their Group Revenue, without requiring those end users to buy an
additional license for the covered use. Coverage may include hosted services
or installations on a customer's devices or cloud account if expressly included.
6.2 Permission for customer installations or redistribution must be included
in the provider's Separate Agreement. Sections 3 and 5 apply unless that
agreement expressly substitutes another acceptance, customer-registration and
reporting arrangement. Commercial payment or provider coverage alone does not
waive recipient registration. Employees acting for the same registered customer
entity need not each obtain a separate commercial license.
6.3 The provider must communicate the scope and duration of coverage to affected
customers. When coverage expires or ends, continued Commercial Use requires
another applicable license or independent eligibility under section 4, except
to the extent the Separate Agreement preserves customer rights. Independent
deployment or redistribution outside the covered product or service is not
covered merely because the customer once received a licensed product.
6.4 A provider relying only on this community License does not confer an
exemption from the revenue threshold on its customers that receive or operate
Model Materials. For example, a company with US$5 million in Group Revenue needs
its own commercial authorization to run the Model locally in an application
whose provider has only this community License. Bundling the weights or
downloading them separately makes no difference.
6.5 A provider operating its own hosted transcription service must qualify
under section 4 or have a Separate Agreement. A customer that only submits audio
and receives ordinary transcripts or timestamps, without receiving or operating
Model Materials, needs no model license or model registration. Deployment in a
customer's devices or cloud account, or operation by a contractor on that
customer's behalf, remains subject to sections 4 and 6.
7. Outputs and independent materials
This License imposes no conditions on ordinary transcripts or timestamps and
does not claim ownership of them. Rights in recordings and their content remain
with their respective holders. Ordinary outputs, independently developed code
and independently developed models are not Model Materials merely because they
are compatible with, evaluated against or used alongside the Model. Separately
licensed materials and rights obtained independently from their owners are not
relicensed by this document.
8. Rights, platforms and contractual obligations
8.1 The Licensor retains its rights in the Original Model. No trademark right
is granted beyond reasonable identification and required attribution. No support,
updates or future releases are promised under this community License.
8.2 The restrictions on modification, commercial use, access and distribution
are conditions of the permissions granted here and contractual commitments You
make by expressly accepting this License. They do not depend solely on whether
every individual parameter or adaptation is separately protected by copyright.
They do not claim rights over independent third-party property or override rights
and exceptions that applicable law makes non-waivable.
8.3 Hosting platforms may have separate terms and rights needed to operate
their services. This License does not amend those agreements. To the extent
permitted by applicable law, You nevertheless agree not to use a platform's
sharing, copying or download features to bypass the acceptance, redistribution
or modification obligations You have accepted here. Nothing in this section
requires a hosting provider to breach its own agreement with the Licensor.
8.4 If You distribute the Model, or offer a product or service to others that
uses it, You must credit it with the notice "Speech recognition by Kroko" and,
where the medium allows, a link to https://kroko.ai. Place the notice where
users can reasonably find it, such as an about page, documentation, credits or
the website of the product or service. The notice must not suggest that the
Licensor endorses or supports Your product or service. Purely internal use
requires no notice. A Separate Agreement may waive or change this requirement.
9. Duration, breach and termination
9.1 This License takes effect upon completion of section 3 and continues for
the Covered Release unless terminated under this section or a Separate Agreement.
You may end Your use at any time. The Licensor may terminate this License for
a material breach by written notice identifying the breach and allowing 30 days
from receipt to cure it. If the breach remains uncured, termination takes effect
at the end of that period. Where applicable law permits, an intentional material
breach or a breach incapable of cure may be terminated immediately by written
notice stating the reason.
9.2 You must promptly stop unauthorized distribution, training or other
noncompliant activity when You become aware of it. A cure period does not
authorize continued prohibited activity or Commercial Use beyond section 4.
Reinstatement after termination requires written confirmation from the Licensor.
9.3 On termination, stop using and distributing the Model Materials, remove
copies You make available to others and delete copies under Your control,
including prohibited adaptations. Copies required by law or held in inaccessible
routine backups may remain only for that purpose and must not be used or restored
for use without a valid license. Termination does not require deletion of
ordinary transcripts or timestamps.
9.4 A distributor's breach or termination does not automatically terminate
compliant recipients' direct licenses from the Licensor. Accrued claims,
ownership provisions, applicable restrictions on retained copies, liability
provisions and dispute provisions survive to the extent needed to give them
effect. Separately granted rights are governed by their own terms.
9.5 Revoking a platform download permission does not itself terminate an
existing license; termination is governed by this section or a Separate
Agreement. This License does not guarantee ongoing hosting availability.
10. Warranty and liability
10.1 To the extent permitted by applicable law, the Model Materials are provided
as is and as available, without express or implied warranties, including as to
accuracy, fitness for a particular purpose, merchantability or noninfringement.
You are responsible for evaluating suitability for Your use and obtaining rights
needed for Your recordings, inputs and separately licensed components.
10.2 To the extent permitted by applicable law, the Licensor and contributors
are not liable for loss or damage arising from the Model Materials or their use,
including loss of profits, business, data or availability, whether a claim is
based on contract, tort or another legal theory.
10.3 These exclusions do not exclude or limit liability for intentional
misconduct, fraud or gross negligence, or for death, personal injury or other
liability to the extent it cannot lawfully be excluded or limited. Mandatory
consumer rights and other non-waivable rights remain unaffected. A Separate
Agreement may provide different warranties or liability terms for its scope.
11. Governing law, notices and versions
11.1 Bulgarian law governs this License. To the extent a jurisdiction agreement
is legally effective, disputes are subject to the competent courts at the
Licensor's registered seat in Bulgaria. This does not deprive a consumer of
mandatory protections or jurisdiction rights under otherwise applicable law,
or displace other rules that cannot lawfully be excluded by agreement.
11.2 Legal notices may be sent to the licensing email above and, for You, to
the current licensing contact supplied under section 3. A notice must be in a
form the recipient can retain; mandatory rules about delivery remain applicable.
11.3 This License applies to the identified Covered Release and the versions
made from it within its scope. Publishing a later license or changing a website
does not retroactively replace the terms accepted for that release. A changed
acceptance URL or administrative process must not change substantive permissions
or obligations for an existing license without agreement.
11.4 A Separate Agreement controls to the extent it expressly changes this
License for the parties, releases and uses it covers. Otherwise this License
remains applicable. If a provision is unenforceable, the remaining provisions
continue to the extent permitted by law. A failure to enforce a provision on one
occasion is not a waiver for another occasion.
End of License
Log in or Sign Up to review the conditions and access this model content.
Kroko-ASR-XL-EN-Offline
Kroko-ASR-XL-EN-Offline is an English speech recognition model for offline transcription. It has 148 million parameters, a Zipformer encoder with RNN-T decoding, and is packaged as ONNX graphs.
In our 10 November 2026 evaluation on the seven default public test sets of the Open ASR Leaderboard, the model reached 5.97% mean WER and 26,879 RTFx with fast-gpu-asr on an NVIDIA H200. Per-dataset results are in Evaluation.
| Task | English speech recognition, offline transcription |
| Architecture | Zipformer encoder, RNN-T (transducer) decoding |
| Parameters | 148,271,160 (approximately 148.3 million) |
| Audio input | Mono, 16 kHz |
| Package | Encoder, decoder and joiner ONNX graphs with tokens.txt |
| Open-source runtimes | sherpa-onnx (CPU and GPU) · fast-gpu-asr (batched, NVIDIA GPUs) · audio.cpp (coming soon?) |
| Commercial runtime | Kroko inference server (details) |
| License | Kroko Community Model License 2.0, gated access; commercial use permitted below US$1,000,000 Group Revenue |
Quick start
- Submit the access form on this page. This records your acceptance of the license.
- Sign in and download the package:
hf auth login
hf download Banafo/Kroko-ASR-XL-EN-Offline --local-dir kroko-asr-xl-en-offline
- Install sherpa-onnx and a WAV reader:
pip install sherpa-onnx soundfile
- Transcribe a 16 kHz mono recording:
import sherpa_onnx
import soundfile as sf
model_dir = "kroko-asr-xl-en-offline"
recognizer = sherpa_onnx.OfflineRecognizer.from_transducer(
encoder=f"{model_dir}/encoder.onnx",
decoder=f"{model_dir}/decoder.onnx",
joiner=f"{model_dir}/joiner.onnx",
tokens=f"{model_dir}/tokens.txt",
decoding_method="modified_beam_search",
max_active_paths=4,
num_threads=4,
)
audio, sample_rate = sf.read("sample.wav", dtype="float32", always_2d=True)
audio = audio[:, 0] # first channel
stream = recognizer.create_stream()
stream.accept_waveform(sample_rate, audio)
recognizer.decode_stream(stream)
print(stream.result.text)
max_active_paths=4 is the beam size. To run on an NVIDIA GPU, install a CUDA build of sherpa-onnx (see its installation guide) and pass provider="cuda".
To try the model without any setup, Zoiper Scribe is a free desktop app for Windows, macOS and Linux that uses it for dictation and file transcription on your own computer.
Files
| File | Purpose |
|---|---|
encoder.onnx |
Zipformer encoder |
decoder.onnx |
RNN-T prediction network |
joiner.onnx |
RNN-T joiner |
tokens.txt |
Token table that maps model outputs to text |
Download all four files together, including any ONNX external-data files stored next to the graphs, and keep their filenames and relative paths intact.
The Open ASR evaluation used RNN-T modified beam search with beam size 6, using mono 16 kHz audio.
Evaluation
Results from our NVIDIA H200 evaluation on the public Open ASR Leaderboard test sets, 10 November 2026:
| Test sets | Mean WER (%) ↓ | Pooled RTFx ↑ |
|---|---|---|
| Seven default public test sets | 5.9684 | 26,879 |
| All ten public test sets | 5.1262 | 26,601 |
Throughput was measured with fast-gpu-asr. Pooled RTFx is total audio duration divided by total timed inference. Figures for the commercial inference server are under Commercial licensing.
Accuracy. RNN-T WER at beam size 6 on all ten public Open ASR test sets. The first seven are the leaderboard's default selection. The last three are reported for completeness.
| Test set | WER (%) ↓ |
|---|---|
| AMI-Cleaned | 7.6808 |
| Earnings22-Cleaned-AA-chunked | 6.4497 |
| GigaSpeech-Cleaned | 7.6063 |
| SPGISpeech | 2.9853 |
| URGENT24 | 10.5635 |
| Voice Arena Monsoon, English–India | 4.0803 |
| VoxPopuli-AA-Cleaned | 2.4127 |
| Average, seven default sets | 5.9684 |
| LibriSpeech Clean | 1.2962 |
| LibriSpeech Other | 3.1018 |
| URGENT24 Clean | 5.0849 |
| Average, all ten sets | 5.1262 |
Averages are unweighted means of the per-set WERs. The evaluation used fast-gpu-asr v0.2.0 with TensorRT FP16, RNN-T modified beam search at beam size 6, a batch size of 256, and three workers. Audio that needed resampling was converted to 16 kHz using float32 polyphase resampling.
Accuracy versus speed. Compared with downloadable models in the 10 November 2026 leaderboard snapshot, this measurement is on the Pareto frontier. Of the 48 downloadable models in that snapshot with a published RTFx, at least 44 are both less accurate and slower.
The five models that no other model beats on both accuracy and speed (the Pareto frontier):
| Model | Mean WER, seven default public sets (%) ↓ | RTFx ↑ |
|---|---|---|
| nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B | 5.51 | 473 |
| Qwen/Qwen3-ASR-1.7B-hf | 5.75 | 836 |
| Kroko-ASR-XL-EN-Offline | 5.97 | 26,879 |
| futo-org/asr4all-m | 10.32 | 28,913 |
| futo-org/asr4all-s | 12.69 | 40,786 |
License summary
This repository is gated. To download the model files, sign in to Hugging Face and submit the access form on this page. Submitting the form is the Official Acceptance Process of the Kroko Community Model License 2.0: it records your acceptance with the licensor, together with the licensee's name, contact details and country, your Hugging Face username and email address, and the time of acceptance. Accept on your own behalf or for one legal entity that you are authorized to bind. Each legal entity registers separately, and an acceptance for one entity does not cover its affiliates.
Permitted use. The community license permits inference, evaluation, format conversion, post-training quantization and supported inference optimizations.
Commercial use. Commercial use is permitted while the licensee's Group Revenue is below US$1,000,000. Group Revenue is total worldwide revenue over the preceding 12 months, consolidated across the licensee's corporate group. Commercial use includes internal business production and products offered free of charge. At US$1,000,000 or more, a separate written agreement with the licensor is required before starting or continuing commercial use, unless a licensed provider's agreement covers that use. There is no grace period for crossing the threshold. Organizations of any size may carry out non-production evaluation after accepting the license.
Redistribution. The original model and permitted conversions may be redistributed only when each recipient has completed this acceptance before receiving the files and the distributor has confirmation of it. The simplest compliant route is to direct recipients to this gated repository. For any other delivery, obtain confirmation of the recipient's acceptance from the licensor first. Do not place the model files in an unrestricted public download, repository, package or container.
Applications. An application may ship the model to its end users, bundled or downloaded by the app, if each user accepts the license inside the application before the model is first made available. Those users do not need to register here. They may run the model only within that application, and the revenue threshold applies to their own commercial use.
Attribution. Products and services that use the model, and redistributed copies, must carry the notice "Speech recognition by Kroko" with a link to https://kroko.ai where users can reasonably find it. Purely internal use needs no notice, and a commercial agreement can waive it.
Hosted services. A customer who only submits audio to a hosted transcription service and receives transcripts or timestamps, without receiving or running the model files, needs no license or registration. The operator of the service must itself qualify under the license or hold a separate agreement.
Not permitted without a separate agreement. Fine-tuning, learned adapters, additional or replacement learned prediction heads, and model merging require express written permission, including when the supplied encoder remains frozen.
This summary is provided for convenience. The complete license controls.
Commercial licensing
Commercial licensing options include:
- Optimized inference engine: deployment options for throughput, latency and VRAM efficiency.
- CTC-aligned timestamps: alignment of recognized text to the audio timeline.
- Overlap decoding: decoding and reconciliation across overlapping audio windows.
- Fine-tuning: permission to adapt the model to customer requirements under express commercial terms.
- Additional languages: access to language offerings covered by the commercial agreement.
Use the licensing contact identified in the license for available features, language coverage and deployment terms. Fine-tuning rights must be expressly included in the agreement. Feature availability is governed by the commercial agreement.
Inference server performance
The Kroko inference server is the commercial runtime for this model. Measured on a single NVIDIA RTX PRO 6000 Blackwell Max-Q, using pipelined batched inference. Peak throughput is 27,713 RTFx (batch size 256, 5-second clips); at batch size 1, p99 latency is between 6.7 and 10.7 ms.
Each cell shows throughput (RTFx, higher is better) / p99 latency per batch cycle (lower is better), by batch size and clip duration:
| Batch size | 5 s clips | 10 s clips | 20 s clips | 30 s clips |
|---|---|---|---|---|
| 1 | 913 / 6.7 ms | 1,682 / 8.4 ms | 2,714 / 8.0 ms | 3,158 / 10.7 ms |
| 8 | 6,286 / 8.6 ms | 10,958 / 8.1 ms | 12,590 / 15.2 ms | 13,283 / 20.4 ms |
| 16 | 10,445 / 8.6 ms | 16,845 / 11.2 ms | 17,575 / 21.6 ms | 17,013 / 32.3 ms |
| 32 | 15,374 / 12.1 ms | 22,210 / 17.4 ms | 21,241 / 35.8 ms | 19,637 / 56.6 ms |
| 48 | 18,588 / 17.5 ms | 23,160 / 25.5 ms | 21,362 / 50.3 ms | 19,527 / 81.4 ms |
| 64 | 20,292 / 18.6 ms | 24,081 / 32.9 ms | 21,776 / 68.9 ms | 19,696 / 110 ms |
| 96 | 24,028 / 24.2 ms | 26,453 / 44.0 ms | 23,309 / 97.3 ms | 20,574 / 159 ms |
| 128 | 24,704 / 31.8 ms | 25,851 / 58.5 ms | 22,441 / 129 ms | 19,874 / 215 ms |
| 256 | 27,713 / 61.5 ms | 27,231 / 114 ms | 22,784 / 287 ms | 19,889 / 479 ms |
RTFx is the duration of processed audio divided by wall time. These figures do not describe sherpa-onnx or fast-gpu-asr.
Apple M5 Max reference
Reference measurements on Apple M5 Max using 30 seconds of audio:
| Configuration | Processing time ↓ | RTFx ↑ |
|---|---|---|
| CPU, 4 encoder threads | 152.5 ms | 196.7× |
| MLX GPU encoder and feature extraction, CPU decoder | 24.59 ms | 1,220× |
Limitations
Recognition errors vary with acoustics, accents, recording conditions and speaking style. Evaluate representative audio for the intended application, and review transcripts where errors could materially affect decisions.