We're excited to announce our first reseller partnership with Line21. We're also pleased to share that our first commercial prototype is already underway: a proof of concept for a 24/7 Spanish → English voice dubbing solution for a Spanish news channel. This marks the beginning of Babelbit's go-to-market strategy. Rather than building a large services organization ourselves, we'll work with specialist resellers who have deep expertise in specific industries and customer segments, bringing Babelbit's technology to the markets they know best. This is the first step, and we look forward to welcoming many more reseller partners in the future. $TAO

Happy New Year! It's our first day back, and we took a unanimous decision to bring forward our first end-to-end speech translation system, starting work on the deployment architecture ... immediately. This will be a lot more fun to watch, because as well as a huge step towards creating a SOTA product, it will allow us to do live demos of the miners' latest achievements improving latency and accuracy week on week. Watch this space... $TAO #SN59

See below the post for what the SimulS2S-LLM team said about their limitations in their paper this year. Note that essential component of our strategy is to use an LLM design designed specifically for speech translation. The prediction gains we are achieving could probably be even greater with the incredible power of something like GPT-5.2, Grok Super Heavy or Gemini 3 Pro Ultra. However it will all be more than negated by the latency of switching modes from speech to text etc. Try a really simple true multi-modal task for yourself "Output audio from this scan of a the musical notation of Twinkle Twinkle Little Star". You'll end up being advised to manage the steps yourself, i.e. it is doable, but only with a multi-phase workflow. An agentic model could probably do it, but still with multiple steps, i.e. adding latency at each step. The job Josh was doing with GANs before joining Babelbit was based on another of my ideas which I thought was intrinsically multi-phase: cleaning noisy speech by recognising what is said and the qualities of the specific speaker and then resynthesising it. I figured the latency would be 250ms - 500ms. Josh got it down to <50ms by a combination of collapsing the phases into a one-shot process, and (the idea which inspired Babelbit) predicting the next chunk of the waveform. Anyway - check out this from the SimulS2S-LLM paper. I found it to be a sad reflection on how R&D is funded. They should have been given the budget to build their own LLM architecture from scratch. Limitations This paper has the following limitations. 1. SimulS2S-LLM relies on off-the-shelf text- based LLMs, meaning that the performance is inherently constrained by the capabilities of the available text-based models. Due to limited computing resources, this paper fo- cuses on using 7B/8B LLMs, as larger models are beyond our computational capacity. Addi- tionally, SimulS2S-LLM is restricted to open- source LLMs and cannot use closed-source models like GPT-4. 2. This paper does not focus on scenarios with extremely low latency (e.g., AL < 1 s in Simul-S2TT), as prior work (Deng and Wood- land, 2024b) indicates these are unsuitable for offline-trained models and compromise trans- lation quality due to the need for re-ordering. Moreover, using LLMs increases the compu- tational load, which leads to higher latency when considering computation time. This is a common challenge faced by the commu- nity, and significant research development is needed to accelerate LLM inference speed. As such, this aspect is beyond the scope of this paper and is left as future work. In ad- dition, since this is the first work to apply LLM to the Simul-S2ST task, we couldn’t find an LLM-based method to compare with SimulS2S-LLM on the Simul-S2ST task. 3. As mentioned in Section 3.4, during the simul- taneous inference of the proposed SimulS2S- LLM, when a new speech chunk is read in, the past keys and values need to be updated before LLM generation. However, according to the analysis in Appendix D, the time consumed by this process should not be significant as it can be performed in parallel. This paper has also not evaluated SimulS2S-LLM on long-form Simul-S2ST due to the lack of data. 4. Due to limitations in training data and com- puting resources, we were unable to train our SimulS2S-LLM as extensively as some foun- dation models like SeamlessStreaming (Bar- rault et al., 2023) which uses 9,300 hours of speech-speech data. However, we conducted comprehensive experiments across three lan- guage pairs and CVSS-C is the most widely used speech-to-speech translation data set,beven though the individual language pairs have only between 69.5 and 174 hours of speech-to-speech data. In addition, if more speech-to-speech translation data is used, the speech generation performance of SimulS2S- LLM can be expected to greatly improve. $TAO #bittensor #sn59 @RvCrypto

We are about to move to the next stage of our prediction development, by adding both multilingual challenges (adding German and Chinese), and by adding an additional fundamental innovation which will revolutionise speech translation latency. In order to take advantage of a prediction, the system needs to know which predictions are good and which are bad, and to measure that confidence level. The multi-lingual challenges will have three separate dialogues (note these are not translations of each other). See image for the new JSON layout. Obviously we are working on new releases for validators to score these as there will be some differences in our incentives as the results will be much richer, e.g. - miners could excel in Chinese but not in German - miners could excel in confidence calculation while taking a hit on accuracy There's a lot of work to be done, and we are collaborating with @const_reborn's team at Affine #SN120, the Data Universe team #SN13 at Macrocosmos, and are working on new security models with Chutes #SN64 and have started working with my namesake @the_mattyk at Manifold with some of Targon's #SN4 VM encryption ideas and the Intel and Nvidia hardware attestation system, which (when we start adding real-time contests which measure actual speed), will ensure that miners and validators are using like-for-like hardware. Exciting times. $TAO #SN59 #babelbit

Bon weekend - and end of another great week. First priority is that we have been building a centralised DB for all the validators' logs, which will enable us to monitor consensus more closely and identify the reasons for any divergence. In parallel we have been preparing the next changes to our competition. The first one will go live in a week - multiple languages. There are a couple of different benefits of this: 1. Multilingual challenge - optimise a single model to cope with more than one language 2. Separate challenges for each language - e.g. flesh out whether optimising for German is different from English Finally Josh, our Chief Scientist, has been working on a model to calculate and/or categorised different levels of confidence. This is crucial and will be added to the competition later in the month. This is an innovation which no other translation technology has. Even if there is some use of next word prediction, that fails to exploit to the ability of an LLM to predict much further. Some utterances are intrinsically more predictable than others and the confidence model tells us which. So now we need the community to start training it.

Last week was a crazily busy week, focusing on multiple collaborations within the Bittensor community. Firstly - special thanks to Jacob @const_reborn for reaching out to us. It is great to find ourselves collaborating with the team at Affine, as we both work towards the next phases of our respective subnets. Back in Sept, I was sending a link to the Affine repo to everyone I could. It looked tailor-made for our needs. Just reading the words sybil-proof, decoy-proof, copy-proof, overfitting-proof reassured me that we wouldn't have to solve every problem ourselves. We are also collaborating with several other subnets... @chutes_ai are hosting models for us, our miners and validators @MacrocosmosAI who are looking into creating bespoke data pipelines with real world dialogue in any language we choose. Finally - a new one for us: - @TargonCompute have just talked us through their decentralised encrypted VMs, and their built-in GPU attestation to verify hardware specs e.g. ensure that specific Nvidia or Intel GPUs are in use, when people say they are. This week is proving to be just as hectic, but more focused on delivery of the next subnet upgrade (rolling out in two stages over the next few days). These changes will help guarantee validator consensus and reduce cheating, making the competition fairer for miners. Full details after our sprint review tomorrow ...

Excellent week at Babelbit SN59. We are developing a new challenge. There is no point in making the best possible predictions without knowing which ones are good. So every prediction needs an associated confidence quotient. This is not a trivial calculation and the best way to perform it, is to train an LLM, i.e. an ideal scenario for a mining challenge. Think about this scale from zero to one: 1: Near zero predictability - 1.1: Private knowledge: "My favourite records are ... " 1.2: Arbitrary lists: "Today's attendees are ... " 2: Near total predictability - 2.1: Quotations: "Our father who ... " 2.2: Nominal phrases: "UNESCO - United Nations Educational Scientific and Cultural Organization So, we are lining up a new base script for release in December, which will have a basic model, and some suggested approaches to future iteration. Watch this space for details of experiments from our R&D work stream. #tao #bittensor #MachineLearning #LLM