Tether Releases Voice Recognition Model for Local Use on Mobile and PC

Tether's QVAC launched TranscribePsy in October 2026, a voice recognition model that can run locally on mobile phones and laptops. It has 254 million parameters, with the laptop version supporting 25 European languages and the mobile version supporting 8. The model weights are open under the Apache 2.0 license, allowing users to download and perform voice transcription on their devices.

Local voice transcription reduces the need to upload raw audio.

For many, the most direct experience of voice recognition is pressing the record button and later receiving a text. TranscribePsy attempts to bring the recognition process back to the device itself, allowing voice to remain local rather than defaulting to the cloud for transcription. It is noteworthy that edge AI is beginning to expand from chat and image generation to basic functions in everyday devices; its applicability must also be judged based on language support, device performance, and data processing chains.

Once such functions enter the workplace, voice transcription will shift from being an "occasional tool" to a data entry point. Interviews with journalists, lawyer meetings, doctor dictations, and internal corporate discussions may first be recorded and then transformed into searchable text. Moving processing from the cloud to the local device will change the systems through which data passes and who maintains the devices and files.

1. Model Brings Voice Transcription to the Device

Cloud-based voice recognition typically requires sending audio to a service provider's server, which then returns the text. Edge models perform inference on the phone or computer, making network connectivity no longer a prerequisite for completing transcription. Offline meeting notes, personal voice memos, and environments with unstable internet conditions are all scenarios where this method could be effective.

TranscribePsy supports multiple languages through a base model combined with language adapters. The version disclosed by QVAC covers 25 European languages on laptops and 8 on mobile; Chinese is currently not on the support list. The mobile version has tighter memory, chip, and battery constraints, resulting in a smaller range of supported languages compared to the laptop version. For Chinese users, it currently showcases the edge voice technology route, and Chinese transcription still requires other solutions.

The language coverage differs between the computer and mobile versions, and the current version does not support Chinese.

The publisher states that the model weights are licensed under Apache 2.0, allowing developers to download and deploy them on devices. This release opens up the model weights and related evaluation configurations, but training code and training data are not included in the release package. Developers must still verify the licensing scope of the model, inference tools, and application components during integration.

QVAC explains that this model is derived from a larger voice recognition model, using the original model as a "teacher" for continued training. It employs the FastConformer-TDT architecture with approximately 254 million parameters, with the team claiming training covered about 564,000 hours across 25 languages. For product developers, the significance of the parameter count is that it is easier to fit into memory and perform calculations on mobile; fewer parameters do not automatically guarantee accuracy for every language, as training samples, accents, and device implementations are equally important.

The desktop version's multilingual design has another detail: English is processed by the base weights, while the other 24 languages come with adapters; when the user does not specify a language, the system first determines the language type and then selects the corresponding path. The mobile version is reduced to eight languages, covering English, Dutch, French, German, Italian, Polish, Portuguese, and Spanish. If a company plans to deploy it in a multinational team, it should test based on the actual working language and not interpret "support for 25 languages" as meaning every phone can use all languages.

This version handles a completed segment of speech or a full recording and does not yet provide real-time streaming transcription. Quickly generating minutes after a meeting differs from displaying subtitles in real-time during a phone call, as the technical requirements are different. QVAC states that streaming functionality is still in development; companies needing real-time dictation should choose products based on current capabilities and not treat the roadmap as delivered functionality.

2. Offline Operation Reduces Audio Transmission

By placing inference locally, the most apparent change is that the audio processing path is shortened: devices can directly output text after recording, reducing the need to upload raw audio to third-party servers. This has practical value in scenarios involving customer communication, internal meetings, or personal privacy, and it also reduces reliance on continuous network connectivity.

The privacy effect also depends on logs, backups, and the application's own data flow.

However, whether the voice leaves the device still depends on how the complete application is designed. Software may synchronize transcription results to the cloud, send crash logs or usage statistics to the service provider, or call cloud services to complete summaries, translations, or retrievals after recognition. Companies should verify microphone permissions, audio cache locations, log contents, automatic backups, and model update channels during deployment. Edge inference changes the processing location, and the privacy effect is ultimately determined by the product's data flow.

This also makes device management a new workload. Organizations need to arrange for model version updates, malware protection, device loss handling, and local audio file deletion policies. When dealing with regulated data, they should also assess whether employee notification, customer consent, access control, and retention period management are necessary based on specific business needs. The model can provide privacy-friendly foundational capabilities, but organizations still need to integrate it into existing data governance processes.

There is an easily overlooked scenario here. An employee records a client meeting using a personal phone; although the audio is not uploaded to the transcription service, the phone has enabled system photo backup, and the transcription text automatically syncs to personal cloud storage. If organizations only check whether the model runs locally, they may overlook the subsequent data flow. A more prudent approach is to use managed devices or managed application containers, limiting file storage locations, encryption methods, backup settings, and export permissions, and informing employees when they must stop recording.

The device side will also change the focus of information security teams. In the cloud model, companies can centrally manage service provider accounts and API calls; in the local model, they also need to know which terminals have installed the model, which version they are using, whether there are known vulnerabilities, and how to remotely lock lost devices. In scenarios where recording content is particularly sensitive, audio and text can be stored in directories controlled by the enterprise, with expiration cleaning and access auditing set up. This is how the advantage of "audio not being recognized through the cloud" can truly be retained.

From a legal perspective, whether to record, whether transcription is allowed, and how long to retain it still depend on the specific scenario and applicable jurisdiction. For example, the purposes of personal information used in client interviews and internal training are different; conversations involving trade secrets must also consider confidentiality agreements and access scope. Edge recognition can reduce one external recipient but does not replace the enterprise's decision on the legal basis for recording. In practice, recording notifications, purposes, retention periods, and internal access permissions should be written into meeting or interview processes, not just in technical solutions.

3. Vendor Test Results Need to Be Contextualized

QVAC disclosed benchmark tests indicating that TranscribePsy achieved good recognition results in its selected comparisons of similarly sized models and tested running speeds on six phones. These results come from the publisher's testing environment and can be used to understand the project team's testing goals and device range. Different devices, accents, noise levels, recording distances, and specialized vocabulary will still affect actual performance; the publicly available performance evidence at this stage mainly comes from vendor testing.

The publisher provided some verifiable numbers. In average tests across eight languages in the multilingual LibriSpeech dataset, it reported about 6.58 errors per 100 words, while the numbers for selected comparable models were around 11 or more. The testing metric is called "word error rate," which counts omitted words, incorrect words, and extra words; the lower the number, the closer the recognition is to the reference text in that test set. This data indicates that the team achieved good results on public benchmarks, but if a meeting is filled with specialized abbreviations, names, or dialects, the actual revision time will still depend on the organization's own samples.

Performance testing should also clarify devices and metrics. QVAC ran the same 40-minute, eight-language audio samples on six phones and published processing times on iPhone, Samsung, and Pixel devices. The iPhone's Core ML version can utilize the neural network processor, while the Android device's GGUF version mainly relies on the CPU, resulting in significant speed differences for the same model across different machines. If a product promises users "transcription in seconds for an hour of recording," it should specify the phone model, whether the model is pre-loaded, recording length, and testing conditions.

The effectiveness of voice recognition is closely related to usage conditions. The same recording in a quiet office and an open meeting room may yield different results; proprietary terms in legal, medical, and financial contexts can also significantly affect recognition quality. Organizations should test with their own recording samples before procurement or deployment, measuring word error rates, processing speeds, device resource consumption, and manual correction times separately. The publicly available benchmark results at this stage mainly come from the publisher, and independent third-party retests are still worth observing.

The operating costs of edge models will also fall on specific devices. Model downloads require storage space, the inference process will consume memory and computing resources, and there will be power consumption and heat generation. Computers may be more suitable for batch processing longer recordings, while mobile devices are better for short, spontaneous recordings; the experience on different devices needs to be evaluated separately.

If companies plan to use it for client interviews or internal meetings, they can first conduct small-scale tests with different accents, meeting distances, background noise, and industry vocabulary while also recording transcription proofreading times. When personal information is involved, they should confirm recording purposes, access permissions, retention periods, and handling plans for lost devices. For specialized scenarios, transcription results still require manual review, especially for key content such as amounts, names, medical terms, and contractual obligations.

Running on the device will transfer some data control to terminal management. Companies need to decide which devices can download the model, how model versions are updated, whether users can export text, when offline files are deleted, and whether application logs include recording snippets. Incorporating these issues into procurement evaluations will allow edge deployment to align with the organization's existing data governance rules.

To reduce misjudgments in pilot projects, a small-scale dual-track evaluation can be conducted. Select real but authorized recordings from meetings, interviews, or classes, manually create a proofreading draft, and have the local model and existing cloud services transcribe separately; then compare accuracy rates, the time required for manual organization of an hour of recording, device battery usage, and maintenance costs for the security team. If the edge model performs well in quiet environments but degrades significantly in group discussions, its applicability can initially be limited to personal dictation or short voice memos. This conclusion is more useful than merely looking at a leaderboard.

Once put into actual use, the responsibility for proofreading after transcription must also be addressed. Meeting minutes will enter decision records, interview drafts may become reporting or research materials, and recordings of contract negotiations may even be used for dispute resolution. The recognition system is only responsible for producing text; organizations should mark which paragraphs have been manually verified, who modified names and numbers, and how long the original audio is retained. If unverified transcripts are directly handed over to a summarization model, which then automatically sends emails, a misheard amount could pass through several systems, ultimately appearing as a formal conclusion.

For application developers, open-source weights lower the deployment threshold but do not exempt them from device adaptation work. Recording inputs need to be converted to the format required by the model, long audio must be processed in segments, and punctuation, speaker differentiation, and specialized vocabulary may need to be supplemented by other components. Each additional component introduces another layer of licensing and data processing issues. If a product claims to be "completely offline," it should check whether the initial download, model updates, crash feedback, and account logins still rely on the internet; if companies only require sensitive audio to remain local, non-sensitive update requests can be allowed to connect online, with the key being to clarify the two types of traffic.

Economically, cloud services charge based on usage and bear most of the computational maintenance, while local operation distributes computation across phones and computers. For individuals occasionally transcribing ten minutes of audio, the difference may be minimal; teams processing large amounts of interviews or meetings daily will be more concerned about subscription costs, device wear, and employee proofreading time. Procurement decisions cannot simply compare whether the model is free. The model's openness, deployment software maintenance, terminal upgrades, and technical support are each borne by different entities, and companies need to include these costs in the same table.

Another straightforward product issue is whether ordinary users are willing to wait for the model to download and load. QVAC disclosed that the peak memory on mobile is approximately 530 to 542 MiB, and different systems may impose restrictions on background applications. Older devices may be able to run the model but may be reclaimed by the system when switching applications. Public-facing products should provide minimum device requirements, initial download sizes, offline availability, and handling methods for failures; otherwise, while technically deployable, user experience may still be hindered at the first step.

4. Competitive Edge of Edge Voice Models Lies in Deployment Experience

TranscribePsy demonstrates a clear product direction: transforming voice recognition from a server function into a local capability executable on devices. For developers, open weights mean integration can be centered around mobile phones, computers, smart glasses, or robots; for organizational users, adoption will depend on whether language support, terminal compatibility, recognition quality, and operational solutions meet specific business needs.

Next, it is worth observing support for Chinese, real-time streaming recognition, and independent testing across different devices. The publisher states that real-time streaming recognition is still in development. As edge models enter more hardware, competition will increasingly focus on startup speed, power consumption, accuracy, and the design of application data processing. Ultimately, users care about whether the model can reliably complete work on their devices and language environments, and whether the product clearly explains the destinations of recordings, transcribed texts, and logs.

Tether's layout in local AI also follows a product logic: voice first becomes text, and other edge models then have the opportunity to perform translation, retrieval, or Q&A. QVAC also introduced related model concepts for vision and translation, but each combination function has its own data flow and quality requirements. The release of the voice model proves that some foundational capabilities can already run on personal devices; whether subsequent products can form a reliable experience still depends on language coverage, application design, and whether users are genuinely willing to manage data locally.

For Chinese readers, the most immediate limitation remains language. Chinese is not on the support list for this release, so it cannot temporarily replace local Chinese meeting transcription tools. However, this model showcases a change worth tracking: as mobile phones gain sufficient recognition capabilities, whether voice data must be uploaded to the cloud will gradually become a question that product design can actively choose to address.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.