With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
BabelBit is a specialized Bittensor subnet designed for real-time, low-latency speech translation across languages. Its core purpose is to eliminate the usual delays in live speech translation by using predictive AI techniques. In traditional simultaneous interpreting (and conventional machine translation), a translator often must wait for a speaker’s sentence or clause to finish (e.g. waiting for a final verb in German) before producing an accurate translation. BabelBit tackles this problem by reframing translation as a “predictive utterance completion” task – essentially, guessing the rest of a speaker’s sentence on the fly so that translation can start before the speaker finishes speaking. By harnessing large language models (LLMs) trained for next-word prediction and global coherence, BabelBit tries to anticipate what the speaker will say next and commit to a translation as soon as it can be adequately predicted.
In practical terms, BabelBit aims to enable speech-to-speech translation “at the speed of thought,” maintaining a natural conversational flow without the awkward pauses traditionally associated with live translation. A speaker could talk in one language, and listeners would hear the translation almost immediately in their own language, with minimal lag. Importantly, BabelBit’s approach isn’t just about speed – it also strives to preserve the meaning and nuance of the original speech, including emotional tone and intent. The subnet’s architecture is being optimized to retain elements like phrasing and sentiment so that the translated speech carries the same emotional impact as the source. The primary use cases envisioned for BabelBit include multilingual live meetings, international conferences, real-time subtitles for broadcasts, and any scenario where people speaking different languages want to communicate fluidly without waiting on slow translation systems. In essence, BabelBit’s technology could power live translators for video calls or on-site interpretation systems, allowing cross-language conversations to happen almost as seamlessly as same-language conversations. The project’s ambition is to deliver translation quality on par with the best current systems, but with near-instantaneous delivery – in fact, the team has set its sights on surpassing the capabilities of services like Google Translate by 2026.
BabelBit is a specialized Bittensor subnet designed for real-time, low-latency speech translation across languages. Its core purpose is to eliminate the usual delays in live speech translation by using predictive AI techniques. In traditional simultaneous interpreting (and conventional machine translation), a translator often must wait for a speaker’s sentence or clause to finish (e.g. waiting for a final verb in German) before producing an accurate translation. BabelBit tackles this problem by reframing translation as a “predictive utterance completion” task – essentially, guessing the rest of a speaker’s sentence on the fly so that translation can start before the speaker finishes speaking. By harnessing large language models (LLMs) trained for next-word prediction and global coherence, BabelBit tries to anticipate what the speaker will say next and commit to a translation as soon as it can be adequately predicted.
In practical terms, BabelBit aims to enable speech-to-speech translation “at the speed of thought,” maintaining a natural conversational flow without the awkward pauses traditionally associated with live translation. A speaker could talk in one language, and listeners would hear the translation almost immediately in their own language, with minimal lag. Importantly, BabelBit’s approach isn’t just about speed – it also strives to preserve the meaning and nuance of the original speech, including emotional tone and intent. The subnet’s architecture is being optimized to retain elements like phrasing and sentiment so that the translated speech carries the same emotional impact as the source. The primary use cases envisioned for BabelBit include multilingual live meetings, international conferences, real-time subtitles for broadcasts, and any scenario where people speaking different languages want to communicate fluidly without waiting on slow translation systems. In essence, BabelBit’s technology could power live translators for video calls or on-site interpretation systems, allowing cross-language conversations to happen almost as seamlessly as same-language conversations. The project’s ambition is to deliver translation quality on par with the best current systems, but with near-instantaneous delivery – in fact, the team has set its sights on surpassing the capabilities of services like Google Translate by 2026.
BabelBit is both an AI-powered translation service and a decentralized network platform. On the product side, the team is building a speech-to-speech translation application that can be offered as a SaaS product for live communication. According to the BabelBit whitepaper, the solution will be provided as a cloud-based service for meetings and collaboration platforms, with an option for on‑premises deployment for users in security-sensitive sectors. This means end-users might experience BabelBit as an integration in video conferencing software or as a standalone translator app that delivers instantaneous multilingual subtitles or spoken translations in real time. A distinctive feature of the product design is its two-stream output: BabelBit plans to produce a fast, real-time translated speech stream for immediate conversational use, plus a second, accuracy-optimized translation as a “record of truth” after a brief delay. In other words, listeners get a quick translation to keep the dialogue flowing, and a refined translation shortly after to ensure nothing was lost or mistranslated – a balance between speed and accuracy.
Under the hood, BabelBit’s build is structured as a Bittensor subnet – meaning it leverages Bittensor’s decentralized AI network infrastructure. The BabelBit subnet (Subnet 59) consists of miners (AI model servers) and validators distributed across the network. The miners are participant-contributed models (e.g. translation AI models) that generate translations/predictions, while validators are nodes that evaluate those translations and score the miners. This incentivized design allows many independent developers and researchers to contribute their own translation models (miners) to BabelBit; the best-performing models (those that translate quickly and accurately) earn higher reputation and rewards in the network’s token system (TAO/𝛼 tokens). Meanwhile, the validators use a set of metrics to judge the quality of translations – BabelBit defines custom adequacy metrics (like EATP, Lead, ACS, etc.) and is even developing an LLM-based judge to score how well a predicted translation conveys the meaning of the original speech. This ensures that miners aren’t just fast but also semantically correct. If a miner’s prediction conveys the same meaning as the true sentence (even with different wording), it can score highly. The network continually adjusts miners’ weights based on these evaluations, creating a competitive “evolutionary” improvement loop for the translation models.
From a technology integration standpoint, BabelBit’s build brings together several components and tools:
Speech and Language AI Modules: BabelBit combines automatic speech recognition (ASR), machine translation (MT), and text-to-speech (TTS) into a pipeline for speech-to-speech translation. (In early phases, some of these may be simplified – for instance, initial challenges use text input/output – but the end goal is a full speech-in, speech-out system.)
Predictive LLM Engine: At the heart is an LLM-based predictor that performs utterance completion. This engine uses the preceding words in a speaker’s sentence to predict the remainder of the utterance, enabling “simultaneous” translation without waiting for the last word. The LLM’s global language modeling allows it to anticipate, for example, what a verb at the end of a German sentence will likely be, based on context. BabelBit is actively researching prompt strategies and fine-tuning to push this predictive capability to its limits.
Chutes Platform: Deployment of models is handled via Chutes (a decentralized model hosting and serving platform in the Bittensor ecosystem). Developers package their translation models (as Dockerized “chutes”), and the BabelBit validators call these to get inference results. Chutes manages the execution of models, allowing the BabelBit network to query any miner’s model through a uniform API.
Bittensor Blockchain Integration: The subnet uses Bittensor’s substrate-based chain for registering miners, validators, and tracking reputations/stakes. Miners must register their hotkeys and hold stake (𝛼) on Subnet 59, and they earn rewards in proportion to the value they provide (as determined by validator scoring). The Bittensor protocol handles the incentive mechanism, so BabelBit doesn’t need its own payment system – it inherits Bittensor’s token economics for rewarding AI work.
Data and Metrics Store: BabelBit’s validators maintain a Postgres database to persist performance data – for example, storing raw translation logs and scores for each miner’s predictions. This data helps in analyzing improvements over time and can also feed back into model training. The system logs can optionally be shipped to cloud storage (S3) for further analysis.
Supporting Decentralized Services: Committed to a Web3 ethos, BabelBit intends to use other decentralized infrastructure where possible. The team mentions leveraging tools like Hippius and Macrocosmos Gravity (services in the Bittensor/Opentensor ecosystem) for things like distributed compute or data routing. In short, the whole build is meant to run off centralized servers – the models, data, and even the hosting of the service are spread across a network of contributors.
All these components culminate in a product that offers instantaneous multilingual communication. For end users, BabelBit would likely present as an app or API where you speak in one language and hear another language in near real-time. The heavy lifting – the predictive translation computation – is done by the swarm of miners behind the scenes on the BabelBit subnet. This decentralized approach not only crowdsources innovation (anyone can try to plug in a better model), but also aligns incentives: the faster and more accurate a model can translate, the more reward it earns. Ultimately, BabelBit’s product is a next-generation translator that marries cutting-edge AI with blockchain-based decentralization. It aims to make language barriers virtually disappear in live conversations, delivering translations with “speech-in, speech-out” immediacy and human-like quality of understanding.
BabelBit is both an AI-powered translation service and a decentralized network platform. On the product side, the team is building a speech-to-speech translation application that can be offered as a SaaS product for live communication. According to the BabelBit whitepaper, the solution will be provided as a cloud-based service for meetings and collaboration platforms, with an option for on‑premises deployment for users in security-sensitive sectors. This means end-users might experience BabelBit as an integration in video conferencing software or as a standalone translator app that delivers instantaneous multilingual subtitles or spoken translations in real time. A distinctive feature of the product design is its two-stream output: BabelBit plans to produce a fast, real-time translated speech stream for immediate conversational use, plus a second, accuracy-optimized translation as a “record of truth” after a brief delay. In other words, listeners get a quick translation to keep the dialogue flowing, and a refined translation shortly after to ensure nothing was lost or mistranslated – a balance between speed and accuracy.
Under the hood, BabelBit’s build is structured as a Bittensor subnet – meaning it leverages Bittensor’s decentralized AI network infrastructure. The BabelBit subnet (Subnet 59) consists of miners (AI model servers) and validators distributed across the network. The miners are participant-contributed models (e.g. translation AI models) that generate translations/predictions, while validators are nodes that evaluate those translations and score the miners. This incentivized design allows many independent developers and researchers to contribute their own translation models (miners) to BabelBit; the best-performing models (those that translate quickly and accurately) earn higher reputation and rewards in the network’s token system (TAO/𝛼 tokens). Meanwhile, the validators use a set of metrics to judge the quality of translations – BabelBit defines custom adequacy metrics (like EATP, Lead, ACS, etc.) and is even developing an LLM-based judge to score how well a predicted translation conveys the meaning of the original speech. This ensures that miners aren’t just fast but also semantically correct. If a miner’s prediction conveys the same meaning as the true sentence (even with different wording), it can score highly. The network continually adjusts miners’ weights based on these evaluations, creating a competitive “evolutionary” improvement loop for the translation models.
From a technology integration standpoint, BabelBit’s build brings together several components and tools:
Speech and Language AI Modules: BabelBit combines automatic speech recognition (ASR), machine translation (MT), and text-to-speech (TTS) into a pipeline for speech-to-speech translation. (In early phases, some of these may be simplified – for instance, initial challenges use text input/output – but the end goal is a full speech-in, speech-out system.)
Predictive LLM Engine: At the heart is an LLM-based predictor that performs utterance completion. This engine uses the preceding words in a speaker’s sentence to predict the remainder of the utterance, enabling “simultaneous” translation without waiting for the last word. The LLM’s global language modeling allows it to anticipate, for example, what a verb at the end of a German sentence will likely be, based on context. BabelBit is actively researching prompt strategies and fine-tuning to push this predictive capability to its limits.
Chutes Platform: Deployment of models is handled via Chutes (a decentralized model hosting and serving platform in the Bittensor ecosystem). Developers package their translation models (as Dockerized “chutes”), and the BabelBit validators call these to get inference results. Chutes manages the execution of models, allowing the BabelBit network to query any miner’s model through a uniform API.
Bittensor Blockchain Integration: The subnet uses Bittensor’s substrate-based chain for registering miners, validators, and tracking reputations/stakes. Miners must register their hotkeys and hold stake (𝛼) on Subnet 59, and they earn rewards in proportion to the value they provide (as determined by validator scoring). The Bittensor protocol handles the incentive mechanism, so BabelBit doesn’t need its own payment system – it inherits Bittensor’s token economics for rewarding AI work.
Data and Metrics Store: BabelBit’s validators maintain a Postgres database to persist performance data – for example, storing raw translation logs and scores for each miner’s predictions. This data helps in analyzing improvements over time and can also feed back into model training. The system logs can optionally be shipped to cloud storage (S3) for further analysis.
Supporting Decentralized Services: Committed to a Web3 ethos, BabelBit intends to use other decentralized infrastructure where possible. The team mentions leveraging tools like Hippius and Macrocosmos Gravity (services in the Bittensor/Opentensor ecosystem) for things like distributed compute or data routing. In short, the whole build is meant to run off centralized servers – the models, data, and even the hosting of the service are spread across a network of contributors.
All these components culminate in a product that offers instantaneous multilingual communication. For end users, BabelBit would likely present as an app or API where you speak in one language and hear another language in near real-time. The heavy lifting – the predictive translation computation – is done by the swarm of miners behind the scenes on the BabelBit subnet. This decentralized approach not only crowdsources innovation (anyone can try to plug in a better model), but also aligns incentives: the faster and more accurate a model can translate, the more reward it earns. Ultimately, BabelBit’s product is a next-generation translator that marries cutting-edge AI with blockchain-based decentralization. It aims to make language barriers virtually disappear in live conversations, delivering translations with “speech-in, speech-out” immediacy and human-like quality of understanding.
BabelBit is developed by BabelBit Ltd, a UK-based company that is part of the Score Ecosystem. The venture is led by seasoned experts in speech technology and software innovation:
Matthew Karas – Founder of BabelBit. Matthew Karas has over 25 years of experience spanning speech & audio research, large-scale digital media products, and transferring AI ideas from the lab into real-world systems. He serves as Founder of BabelBit and also as Chief Innovation Officer at Score (which runs another Bittensor subnet, SN-44). Karas’s background is rich in both academia and industry. In the mid-1990s he studied at Cambridge under pioneers like Karen Spärck Jones and Tony Robinson, gaining early exposure to neural network approaches for speech recognition. He went on to lead or contribute to several landmark projects in UK media tech – for example, he designed the content management system for BBC News Online in the late 1990s (a system that impressively remained in use until 2025). He also was instrumental in launching ITV Player in 2006-07 (the UK’s first web streaming service with instant-play and integrated ads, launched even before the BBC’s iPlayer). Karas served as launch CTO of FutureLearn (a major MOOC platform) and founded a multimedia startup, Dremedia, which was acquired by Autonomy in 2003. He holds multiple tech patents, including ones in content structuring and speech assessment. Notably, he developed Eloqute, a system for automated pronunciation training in continuous speech (addressing language learning needs). Matthew Karas’s career has been defined by bridging cutting-edge research and practical applications – an approach he brings to BabelBit. His extensive network includes collaborations with BBC executives and academics; for instance, he maintained a 28-year professional relationship with Mike Lynch (founder of Autonomy) focusing on projects in information retrieval and speech tech. At BabelBit, Karas is architecting the subnet’s strategy – specifically, designing the dual-stream translation approach and the incentive mechanisms where AI models compete to minimize latency and maximize accuracy. He began incubating the BabelBit concept while working within Score, and spun it out as a dedicated subnet project in 2025.
Josh Greifer – Chief Scientist. Josh Greifer is the Chief Scientist of BabelBit, bringing deep expertise in low-latency audio engineering and signal processing. Greifer’s career started at the intersection of music technology and computer science. In the late 1980s, he was part of the Steinberg Cubase team, where he helped transform Cubase from a MIDI sequencer on the Atari ST into a full-fledged digital audio workstation on Mac/PC. In doing so, he architected one of the earliest low-latency audio pipelines for personal computers – a crucial innovation for music production that set new standards for responsiveness (legend has it company founder Charlie Steinberg would sleep on Josh’s sofa during intense coding sprints!). After his work in pro audio, Josh ventured into finance technology at Goldman Sachs in the ’90s, and later into biomedical signal processing – notably running a clinical trial on using smartphone sensor data for arrhythmia (atrial fibrillation) detection. In the 2010s, he returned to speech/audio tech as Chief Architect at Speech Engineering Ltd, where he contributed to Eloqute (the same pronunciation training project that Matthew Karas patented). He also worked with the BBC’s R&D department on media platforms like BBC Redux, and with a startup (Neurence) where he helped slash the latency of GAN-based audio processing from over 250ms to under 50ms. This blend of experiences – from real-time audio pipelines to speech AI – makes Greifer ideally suited to BabelBit’s challenges. At BabelBit, Josh Greifer is overseeing the technical R&D to achieve ultra-low latency in the speech-to-speech engine. He is focusing on the “latency-critical audio pipelines and streaming inference” – for example, optimizing how audio input is captured, processed through the LLM-based translator, and output as speech with minimal delay. Greifer is also helping design the evaluation framework that balances latency against accuracy, ensuring that the system knows when to trust a prediction and when to wait for more clarity.
Other Team and Partners: BabelBit’s team is relatively small and focused, but it benefits from a strong support network. The project is backed by the Yuma Accelerator, a Bittensor-focused incubator (associated with Digital Currency Group) that supports promising subnet projects. Yuma’s support was instrumental during BabelBit’s launch in late 2025. Additionally, as noted, BabelBit emerged from the Score ecosystem – Score (Subnet 44) is another AI venture, and BabelBit’s early development was nurtured there. This means BabelBit has access to the broader Score community and resources. Tom Horner is also listed as a team member (role not publicly specified yet), and likely contributes to the technical development alongside Karas and Greifer. Given the backgrounds of the core team, BabelBit is rooted in a deep understanding of both the linguistic challenges (predicting meaning, handling nuances) and the engineering hurdles (real-time streaming, distributed systems) needed to make its vision a reality.
BabelBit is developed by BabelBit Ltd, a UK-based company that is part of the Score Ecosystem. The venture is led by seasoned experts in speech technology and software innovation:
Matthew Karas – Founder of BabelBit. Matthew Karas has over 25 years of experience spanning speech & audio research, large-scale digital media products, and transferring AI ideas from the lab into real-world systems. He serves as Founder of BabelBit and also as Chief Innovation Officer at Score (which runs another Bittensor subnet, SN-44). Karas’s background is rich in both academia and industry. In the mid-1990s he studied at Cambridge under pioneers like Karen Spärck Jones and Tony Robinson, gaining early exposure to neural network approaches for speech recognition. He went on to lead or contribute to several landmark projects in UK media tech – for example, he designed the content management system for BBC News Online in the late 1990s (a system that impressively remained in use until 2025). He also was instrumental in launching ITV Player in 2006-07 (the UK’s first web streaming service with instant-play and integrated ads, launched even before the BBC’s iPlayer). Karas served as launch CTO of FutureLearn (a major MOOC platform) and founded a multimedia startup, Dremedia, which was acquired by Autonomy in 2003. He holds multiple tech patents, including ones in content structuring and speech assessment. Notably, he developed Eloqute, a system for automated pronunciation training in continuous speech (addressing language learning needs). Matthew Karas’s career has been defined by bridging cutting-edge research and practical applications – an approach he brings to BabelBit. His extensive network includes collaborations with BBC executives and academics; for instance, he maintained a 28-year professional relationship with Mike Lynch (founder of Autonomy) focusing on projects in information retrieval and speech tech. At BabelBit, Karas is architecting the subnet’s strategy – specifically, designing the dual-stream translation approach and the incentive mechanisms where AI models compete to minimize latency and maximize accuracy. He began incubating the BabelBit concept while working within Score, and spun it out as a dedicated subnet project in 2025.
Josh Greifer – Chief Scientist. Josh Greifer is the Chief Scientist of BabelBit, bringing deep expertise in low-latency audio engineering and signal processing. Greifer’s career started at the intersection of music technology and computer science. In the late 1980s, he was part of the Steinberg Cubase team, where he helped transform Cubase from a MIDI sequencer on the Atari ST into a full-fledged digital audio workstation on Mac/PC. In doing so, he architected one of the earliest low-latency audio pipelines for personal computers – a crucial innovation for music production that set new standards for responsiveness (legend has it company founder Charlie Steinberg would sleep on Josh’s sofa during intense coding sprints!). After his work in pro audio, Josh ventured into finance technology at Goldman Sachs in the ’90s, and later into biomedical signal processing – notably running a clinical trial on using smartphone sensor data for arrhythmia (atrial fibrillation) detection. In the 2010s, he returned to speech/audio tech as Chief Architect at Speech Engineering Ltd, where he contributed to Eloqute (the same pronunciation training project that Matthew Karas patented). He also worked with the BBC’s R&D department on media platforms like BBC Redux, and with a startup (Neurence) where he helped slash the latency of GAN-based audio processing from over 250ms to under 50ms. This blend of experiences – from real-time audio pipelines to speech AI – makes Greifer ideally suited to BabelBit’s challenges. At BabelBit, Josh Greifer is overseeing the technical R&D to achieve ultra-low latency in the speech-to-speech engine. He is focusing on the “latency-critical audio pipelines and streaming inference” – for example, optimizing how audio input is captured, processed through the LLM-based translator, and output as speech with minimal delay. Greifer is also helping design the evaluation framework that balances latency against accuracy, ensuring that the system knows when to trust a prediction and when to wait for more clarity.
Other Team and Partners: BabelBit’s team is relatively small and focused, but it benefits from a strong support network. The project is backed by the Yuma Accelerator, a Bittensor-focused incubator (associated with Digital Currency Group) that supports promising subnet projects. Yuma’s support was instrumental during BabelBit’s launch in late 2025. Additionally, as noted, BabelBit emerged from the Score ecosystem – Score (Subnet 44) is another AI venture, and BabelBit’s early development was nurtured there. This means BabelBit has access to the broader Score community and resources. Tom Horner is also listed as a team member (role not publicly specified yet), and likely contributes to the technical development alongside Karas and Greifer. Given the backgrounds of the core team, BabelBit is rooted in a deep understanding of both the linguistic challenges (predicting meaning, handling nuances) and the engineering hurdles (real-time streaming, distributed systems) needed to make its vision a reality.
Current Progress (as of late 2025): BabelBit officially launched Subnet 59 on the Bittensor network in October 2025. Upon launch, the project rolled out its initial developer tooling and a proof-of-concept challenge. The first milestone was to tackle the core prediction engine in a simplified form: text-based utterance prediction. In fact, the team’s first community challenge deliberately “doesn’t involve speech and doesn’t involve translation” – instead, it asks participants to build models that can predict the completion of a sentence given only the beginning of it. This task is critical to BabelBit’s mission, as it trains the network to guess the remainder of utterances in a foreign language before they’re fully spoken. To support this, BabelBit released a dataset of 1 million example utterances (formatted in JSON) for miners to practice and test their prediction algorithms. Early participants in the subnet have been improving scripts and models that generate semantically similar continuations of partial input – for example, if the input heard so far is “Hi – how…”, a model might predict the full sentence “Hi – how are you?” which carries essentially the same meaning as the true phrase “how is all going with you?”. By validating that the predicted continuation preserves meaning, even if wording differs, BabelBit can dramatically reduce translation latency without much loss of accuracy.
At the network level, Subnet 59 (BabelBit) has been operational with miners and validators joining the test rounds. The infrastructure (wallets, registration, etc.) for BabelBit is in place, using Bittensor’s standard subsystem. Developers can use the BabelBit CLI (bb command via the uv tool) and the Chutes SDK to deploy their translation models into the subnet. The validators run a Dockerized BabelBit validator node that issues translation challenges (currently text-based dialogues) to miners and scores their responses. This setup has been tested and refined in the weeks following launch. The team has also been actively engaging the community via X (Twitter) and Bittensor’s channels – sharing progress, inviting new contributors, and highlighting that BabelBit’s token (SN59’s α token) started trading with a market cap (FDV around $13.5M as of launch). By Q4 2025, BabelBit successfully proved the viability of its approach in a controlled setting: the subnet can indeed incentivize models to predict and translate text with lower latency than traditional methods.
Near-Term Plans: Going forward into early 2026, BabelBit’s next steps involve progressively adding the real-world complexity that the initial challenge left out. The immediate plan is to integrate actual translation and speech into the loop. According to the BabelBit whitepaper, one upcoming milestone is “live text-to-text translation (e.g. real-time multilingual subtitling)”. This means the system will start handling full translation tasks where the input is text in one language and the output is text in another – but done in a streaming fashion suitable for subtitles. Achieving this will likely involve training or fine-tuning models for simultaneous translation, as well as testing the adequacy metrics on true bilingual data. In parallel, BabelBit is developing a “novel low-latency audio ingest architecture” to handle speech-in, text-out scenarios. This refers to the front-end speech recognition component: capturing a speaker’s voice and converting it to text (or intermediate features) in real time, feeding it into the LLM predictor without introducing bottlenecks. The team is likely exploring efficient streaming ASR models or partial decoding techniques so that speech can be transcribed on the fly. By combining that with their LLM predictor, they aim to output translated text nearly concurrently with the spoken words. This will set the stage for full speech-to-speech translation. A logical short-term goal is to demonstrate a live speech demo – for instance, one person speaks a sentence in Language A, and another person hears the translated sentence in Language B with only a tiny delay.
Mid-Term Plans: Once text-to-text and speech-in/text-out are working with low latency, BabelBit will focus on the final piece: instant speech-out (voice synthesis). The challenge here is to generate natural-sounding speech in the target language on the fly, possibly with emotional nuance and a voice that matches the context. BabelBit’s research roadmap hints at exploring “one-shot speech-to-speech models that bypass intermediate text” (a concept alluding to end-to-end models that directly map input audio to output audio) – though initially they will likely use intermediate text with a TTS engine for reliability. By mid-2026, the goal is to have a fully functional speech-to-speech translation system running on the BabelBit subnet. This includes perfecting the dual-stream approach: the “fast lane” translation that might sacrifice a tiny bit of literal accuracy for speed, and the “slow lane” high-accuracy translation that might arrive a second later for confirmation. Achieving a smooth handoff between these streams is a technical target on the roadmap. We can expect iterative improvements to the adequacy scoring as well – the planned LLM-based judge for translation quality should be implemented to ensure that as miners rush to translate faster, they don’t produce gibberish. The network’s reward mechanism will continuously be tuned so that it incentivizes an optimal balance of speed vs. correctness vs. naturalness in translation.
Long-Term Vision: BabelBit’s ultimate vision is to make real-time translated conversations as natural as untranslated ones. By 2026 and beyond, they aim for “state-of-the-art translation services” delivered via this decentralized network. If successful, BabelBit could power live translators for international business meetings, customer support calls, gaming voice chat, or even wearable translator devices, all while being backed by an open network of AI models rather than a single corporate model. The project’s aspirational target of outperforming incumbent solutions (like Google’s translator offerings) by 2026 sets an aggressive timeline for reaching top-tier translation quality. To get there, the team will likely incorporate the latest research from the field: e.g. advancements in simultaneous translation algorithms, better handling of colloquialisms and code-switching, and transferring the speaker’s tone or emotion into the translated speech. Another future direction is expanding language coverage – BabelBit might start with a few language pairs (such as English/German, English/Spanish) during development, but plans to broaden to many languages as the models and community grow. The Bittensor community aspect also means that, over time, more miners will join with novel model architectures (for example, someone might contribute a cutting-edge speech transformer or an efficient RNN specifically fine-tuned for simultaneous translation). This diversity can accelerate innovation on the subnet.
On the deployment side, as BabelBit matures, we will see the productization phase: offering the service to early adopters and enterprises. The roadmap includes rolling out the SaaS platform with easy integration (APIs or SDKs for platforms like Zoom, Teams, etc.). The team will also pursue pilot programs in sectors like government, healthcare, or defense (where on-premises installations are preferred for confidentiality). These pilots will test BabelBit in mission-critical settings. Given the pedigree of the team, there’s also likely a focus on securing intellectual property (more patents around low-latency translation techniques) and building partnerships (perhaps with device manufacturers for real-time translation earpieces, or with cloud providers to reach more users).
In summary, BabelBit’s roadmap is a stepwise expansion from a theoretical foundation to a full-fledged, real-world translation network. They started by validating the concept of LLM-based predictive translation on text data. Next, they are layering in actual speech and multi-language translation capabilities through 2026. Each stage – from live subtitling to streaming audio – brings them closer to the end goal: a decentralized Babel fish that lets anyone converse across languages instantly. Every indication from publicly available info is that BabelBit is on track with development, guided by an experienced team, community contributors, and support from the wider Bittensor ecosystem. If they continue at this pace, BabelBit could indeed become one of the flagship subnets of Bittensor, showcasing how decentralized AI can solve a longstanding human communication challenge in a novel way.
Current Progress (as of late 2025): BabelBit officially launched Subnet 59 on the Bittensor network in October 2025. Upon launch, the project rolled out its initial developer tooling and a proof-of-concept challenge. The first milestone was to tackle the core prediction engine in a simplified form: text-based utterance prediction. In fact, the team’s first community challenge deliberately “doesn’t involve speech and doesn’t involve translation” – instead, it asks participants to build models that can predict the completion of a sentence given only the beginning of it. This task is critical to BabelBit’s mission, as it trains the network to guess the remainder of utterances in a foreign language before they’re fully spoken. To support this, BabelBit released a dataset of 1 million example utterances (formatted in JSON) for miners to practice and test their prediction algorithms. Early participants in the subnet have been improving scripts and models that generate semantically similar continuations of partial input – for example, if the input heard so far is “Hi – how…”, a model might predict the full sentence “Hi – how are you?” which carries essentially the same meaning as the true phrase “how is all going with you?”. By validating that the predicted continuation preserves meaning, even if wording differs, BabelBit can dramatically reduce translation latency without much loss of accuracy.
At the network level, Subnet 59 (BabelBit) has been operational with miners and validators joining the test rounds. The infrastructure (wallets, registration, etc.) for BabelBit is in place, using Bittensor’s standard subsystem. Developers can use the BabelBit CLI (bb command via the uv tool) and the Chutes SDK to deploy their translation models into the subnet. The validators run a Dockerized BabelBit validator node that issues translation challenges (currently text-based dialogues) to miners and scores their responses. This setup has been tested and refined in the weeks following launch. The team has also been actively engaging the community via X (Twitter) and Bittensor’s channels – sharing progress, inviting new contributors, and highlighting that BabelBit’s token (SN59’s α token) started trading with a market cap (FDV around $13.5M as of launch). By Q4 2025, BabelBit successfully proved the viability of its approach in a controlled setting: the subnet can indeed incentivize models to predict and translate text with lower latency than traditional methods.
Near-Term Plans: Going forward into early 2026, BabelBit’s next steps involve progressively adding the real-world complexity that the initial challenge left out. The immediate plan is to integrate actual translation and speech into the loop. According to the BabelBit whitepaper, one upcoming milestone is “live text-to-text translation (e.g. real-time multilingual subtitling)”. This means the system will start handling full translation tasks where the input is text in one language and the output is text in another – but done in a streaming fashion suitable for subtitles. Achieving this will likely involve training or fine-tuning models for simultaneous translation, as well as testing the adequacy metrics on true bilingual data. In parallel, BabelBit is developing a “novel low-latency audio ingest architecture” to handle speech-in, text-out scenarios. This refers to the front-end speech recognition component: capturing a speaker’s voice and converting it to text (or intermediate features) in real time, feeding it into the LLM predictor without introducing bottlenecks. The team is likely exploring efficient streaming ASR models or partial decoding techniques so that speech can be transcribed on the fly. By combining that with their LLM predictor, they aim to output translated text nearly concurrently with the spoken words. This will set the stage for full speech-to-speech translation. A logical short-term goal is to demonstrate a live speech demo – for instance, one person speaks a sentence in Language A, and another person hears the translated sentence in Language B with only a tiny delay.
Mid-Term Plans: Once text-to-text and speech-in/text-out are working with low latency, BabelBit will focus on the final piece: instant speech-out (voice synthesis). The challenge here is to generate natural-sounding speech in the target language on the fly, possibly with emotional nuance and a voice that matches the context. BabelBit’s research roadmap hints at exploring “one-shot speech-to-speech models that bypass intermediate text” (a concept alluding to end-to-end models that directly map input audio to output audio) – though initially they will likely use intermediate text with a TTS engine for reliability. By mid-2026, the goal is to have a fully functional speech-to-speech translation system running on the BabelBit subnet. This includes perfecting the dual-stream approach: the “fast lane” translation that might sacrifice a tiny bit of literal accuracy for speed, and the “slow lane” high-accuracy translation that might arrive a second later for confirmation. Achieving a smooth handoff between these streams is a technical target on the roadmap. We can expect iterative improvements to the adequacy scoring as well – the planned LLM-based judge for translation quality should be implemented to ensure that as miners rush to translate faster, they don’t produce gibberish. The network’s reward mechanism will continuously be tuned so that it incentivizes an optimal balance of speed vs. correctness vs. naturalness in translation.
Long-Term Vision: BabelBit’s ultimate vision is to make real-time translated conversations as natural as untranslated ones. By 2026 and beyond, they aim for “state-of-the-art translation services” delivered via this decentralized network. If successful, BabelBit could power live translators for international business meetings, customer support calls, gaming voice chat, or even wearable translator devices, all while being backed by an open network of AI models rather than a single corporate model. The project’s aspirational target of outperforming incumbent solutions (like Google’s translator offerings) by 2026 sets an aggressive timeline for reaching top-tier translation quality. To get there, the team will likely incorporate the latest research from the field: e.g. advancements in simultaneous translation algorithms, better handling of colloquialisms and code-switching, and transferring the speaker’s tone or emotion into the translated speech. Another future direction is expanding language coverage – BabelBit might start with a few language pairs (such as English/German, English/Spanish) during development, but plans to broaden to many languages as the models and community grow. The Bittensor community aspect also means that, over time, more miners will join with novel model architectures (for example, someone might contribute a cutting-edge speech transformer or an efficient RNN specifically fine-tuned for simultaneous translation). This diversity can accelerate innovation on the subnet.
On the deployment side, as BabelBit matures, we will see the productization phase: offering the service to early adopters and enterprises. The roadmap includes rolling out the SaaS platform with easy integration (APIs or SDKs for platforms like Zoom, Teams, etc.). The team will also pursue pilot programs in sectors like government, healthcare, or defense (where on-premises installations are preferred for confidentiality). These pilots will test BabelBit in mission-critical settings. Given the pedigree of the team, there’s also likely a focus on securing intellectual property (more patents around low-latency translation techniques) and building partnerships (perhaps with device manufacturers for real-time translation earpieces, or with cloud providers to reach more users).
In summary, BabelBit’s roadmap is a stepwise expansion from a theoretical foundation to a full-fledged, real-world translation network. They started by validating the concept of LLM-based predictive translation on text data. Next, they are layering in actual speech and multi-language translation capabilities through 2026. Each stage – from live subtitling to streaming audio – brings them closer to the end goal: a decentralized Babel fish that lets anyone converse across languages instantly. Every indication from publicly available info is that BabelBit is on track with development, guided by an experienced team, community contributors, and support from the wider Bittensor ecosystem. If they continue at this pace, BabelBit could indeed become one of the flagship subnets of Bittensor, showcasing how decentralized AI can solve a longstanding human communication challenge in a novel way.
From September 2025 at The Realistic Traders quarterly event, this video features a veteran speech technology expert reflecting on three decades of innovation and unveiling plans for a groundbreaking new Bittensor subnet focused on real-time speech translation. Having built his first neural network in 1994 under pioneer Tony Robinson, the Matthew Karas traces his career from early AI work at Sky and the BBC—where he helped develop one of the first speech-based search engines—to founding Dream Media and securing a patent for analyzing continuous speech features. He recounts collaborations with influential figures like Mike Lynch and the tragic loss of two close colleagues, which reignited his mission to push speech AI further. Inspired by breakthroughs in latency reduction and large language models, he reveals his vision to build a Bittensor-powered subnet capable of near-instant, accurate speech-to-speech translation—faster and more precise than Google Translate. With support from the Score team and contributions from other subnets for data and infrastructure, the project aims to democratize speech technology, combining decentralized mining with decades of linguistic expertise to deliver seamless, multilingual communication in real time.
Head over to Yuma's X account at 4pm (UK) / 11am (ET) / 8am (PT)
Matthew will be the guest on Yuma's Subnet Spotlight.
https://x.com/yumagroup
@babelbit @yumagroup @webuildscore
#machinelearning #ai #llms @matthew_karas
Our website has had a major revamp.
- It now has a ticker showing live information. Currently this includes the most recent winning miner and the price of TAO, but we will be expanding this to include links as we add more in-depth content
- the main page now scrolls down to…
Matthew's appearance on the Ventura Labs podcast is now up!
Ventura Labs Ep. 68 – Matthew Karas
Matthew (@matthew_karas), Founder of Babelbit (@babelbit)
Timestamps
01:27 — Introduction
2:09 — Origin of Babelbit Name
3:36 — Matthew's Background in Linguistics
6:05 — Early Career at BBC
8:31 — Noise Elimination Using GANs
10:46 —…
What's the quickest way you have found to explain Bittensor to newbies, outsiders, partners, spouses, children, the elderly, friends, investors, HMRC (or IRS) etc?
The OTF "about page" is pretty thorough, but it is what Medium calls an "11-minute read".
My usual elevator pitch…
Coming soon... (Monday 3rd Nov)
Matthew was a guest on the Ventura Labs podcast this week. I'm not sure whether Grant will have got a word in. Hopefully this will be the first of many, and the rest of the team will get to have their say.
This podcast focused on the origin…
SOTA projects in all aspects of speech and language are converging on Multi-Modal LLMs as a single tool for what used to be separate processes run in series: recognition, translation and synthesis. I thought it was an exaggeration, but the more we dig...
