Spark-Audio Targets General-Purpose Speech Understanding on Domestic Hardware
Spark-Audio 1.0 combines a 0.65B audio encoder with a 30B-A3B MoE language model to handle multilingual speech, translation, audio understanding, and dialogue.
Underlying Paper
Spark-Audio 1.0 Technical Report
This paper introduces Spark-Audio 1.0, a speech foundation model trained on domestically produced computing infrastructure. The model accepts text and speech inputs and generates text outputs. It supports multilingual (99 languages) and multi-dialect (202 dialects) speech recognition, speech translation, audio perception and understanding, and both text- and speech-based dialogue, marking a shift from merely ``hearing clearly'' to genuinely ``understanding'' speech. Architecturally, Spark-Audio 1.0 couples a 0.65B-parameter dense audio encoder with a 30B-A3B Mixture-of-Experts (MoE) large language model (LLM). The audio encoder is pretrained with representation-enhanced BEST-RQ and then refined through multi-task supervised training. The audio--LLM is trained through staged pretraining, followed by post-training that combines supervised fine-tuning with on-policy distillation. In total, the model is trained on 13 million hours of audio data together with large-scale text data. We evaluate Spark-Audio 1.0 on mainstream audio and text benchmarks, as well as on in-house test sets drawn from real-world application scenarios. Despite using substantially less audio training data than Qwen3.5-Omni-Flash, Spark-Audio 1.0 achieves comparable performance on most audio and text tasks, outperforms it in multilingual speech recognition, spoken dialogue, and real-world application scenarios, and approaches the performance of larger closed-source models. Nevertheless, there remains room for improvement in instruction following and complex audio reasoning. Finally, the successful training of Spark-Audio 1.0 on domestic hardware, including Huawei Ascend 910B and Hygon BW1000 chips, demonstrates the feasibility of a domestic technical pathway for building speech foundation models.
Spark-Audio 1.0 is a speech foundation model designed to move beyond transcription toward broader speech understanding. It accepts text and speech inputs and generates text outputs for speech recognition, speech translation, audio perception and understanding, and text- or speech-based dialogue.
Core Contribution
The system couples a 0.65B-parameter dense audio encoder to a 30B-A3B mixture-of-experts language model. The authors report support for 99 languages and 202 dialects, positioning the model as a single foundation model for multilingual and multi-dialect speech tasks rather than a narrowly focused recognizer.
A central contribution is also infrastructural: the model was trained on domestically produced hardware, including Huawei Ascend 910B and Hygon BW1000 chips. The report presents this as evidence that large-scale speech foundation-model development is feasible on that hardware pathway.
Technical Approach
The audio encoder is pretrained with representation-enhanced BEST-RQ and then refined through multi-task supervised training. The audio--language-model stack is trained in stages, followed by post-training that combines supervised fine-tuning with on-policy distillation.
Training uses 13 million hours of audio data alongside large-scale text data. This combination is intended to connect a speech representation with language-model capabilities across recognition, translation, audio understanding, and dialogue.
Results and Analysis
The authors evaluate Spark-Audio on mainstream audio and text benchmarks as well as in-house tests intended to reflect real-world application scenarios. They report performance comparable to Qwen3.5-Omni-Flash on most audio and text tasks despite using substantially less audio training data.
The report further claims stronger results in multilingual speech recognition, spoken dialogue, and real-world application scenarios than Qwen3.5-Omni-Flash. It also states that the model approaches the performance of larger closed-source systems on the evaluated tasks. These findings support the paper's narrower conclusion: the system is presented as a competitive general-purpose speech model, not as an across-the-board leader.
Caveats in Practice
The authors identify instruction following and complex audio reasoning as areas for improvement. Those limitations matter for deployments that require reliable adherence to detailed instructions or robust reasoning over difficult acoustic situations.
The supplied material describes broad evaluation, but it does not establish performance for every deployment setting. Results should therefore be read as benchmark and scenario evidence for a general-purpose speech model, with further validation needed for specialized or high-stakes use cases.
Evidence Box
strongKey Claims
- •A single model supports speech recognition, translation, audio understanding, and dialogue
- •Spark-Audio supports multilingual and multi-dialect speech processing
- •The report finds competitive performance against Qwen3.5-Omni-Flash on most evaluated audio and text tasks
- •The model was trained on Huawei Ascend 910B and Hygon BW1000 hardware
Key Results
- •The system combines a 0.65B dense audio encoder with a 30B-A3B MoE language model
- •Training uses 13 million hours of audio data together with large-scale text data
- •The authors report support for 99 languages and 202 dialects
- •The report claims particular strength in multilingual recognition, spoken dialogue, and real-world application scenarios
Limitations & Caveats
- •Instruction following remains an identified weakness
- •Complex audio reasoning remains an identified weakness
- •The reported comparisons are limited to the paper's benchmark and application evaluations
- •Specialized deployment performance requires additional validation
