An in-cabin assistant improves only as fast as its training data. eBanma aggregates the fleet's daily voice and language volume onto MOI, standardised on a common event schema, with audio, transcripts, embeddings and labels in one engine — so a training corpus is queried rather than hunted for.
eBanma (Banma Network) builds intelligent vehicle operating systems and connected-car solutions, founded in November 2015 with Alibaba Group as its largest shareholder and SAIC as its innovation source. Built on its self-developed AliOS, its cockpit solutions ship with more than ten auto brands including SAIC, FAW and Volkswagen Group.
An in-cabin assistant improves only as fast as its training data. Vehicles on the road generate an enormous daily volume of voice interactions and the language data around them-utterances, transcripts, recognized intents, dialogue context, correction and fallback events-across models, regions, accents and cabin conditions. Making that stream useful for deep-learning training means landing it at scale, keeping it labelled and queryable rather than as loose files on storage, assembling training sets that target specific weaknesses, and doing all of it under the privacy controls that in-cabin audio demands.
eBanma aggregates voice and language data onto one OmniFabric platform. Data streams in continuously from the fleet, is standardized on a common event schema, and is stored alongside its structured context-vehicle model, region, cabin state, dialogue session-with transcripts, embeddings and labels held in the same engine. Teams assemble training sets by query, in SQL and by vector similarity: the utterances where the assistant fell back, the accents a model handles badly, the sessions around one intent. Versioned datasets feed deep-learning training and evaluation, and results are written back as labels so the next iteration targets what actually failed.
Audio, transcripts, embeddings and labels sit in one engine with their vehicle and session context, so the training corpus is searched rather than hunted through directories.
A new experiment begins as a query-this intent, this region, these fallback cases-instead of a data-collection project, so iteration speed is set by the model team rather than the pipeline.
Vehicle model, region, cabin state and dialogue session stay attached to every utterance, which is what makes it possible to isolate where a model actually underperforms.