Artificial intelligence has achieved a new breakthrough in its exploration of speech technology. iFlytek recently officially announced that its flagship speech foundation model, SparkAudio1.0, has fully launched API services on the iFlytek Open Platform, and developers and enterprise users can now directly integrate and call the service.

Farewell to Traditional Cascaded Solutions: Using a Native Speech Large Model to Directly Understand the Sounds of the World

For a long time, the industry generally adopted a “cascaded solution” for audio tasks: first converting sound into text through speech recognition, and then passing the text to a large model for understanding. However, this traditional serial approach has two shortcomings that cannot be ignored:

Severe information loss: Pure text can only record “what was said,” but completely misses key dimensions inherent in sound, such as tone, emotional fluctuations, speaker characteristics, and background environmental sounds, resulting in incomplete judgments by the model in complex scenarios.

image.png

Fragmented pipeline and latency: Splitting multiple capabilities—such as speech transcription, voiceprint recognition, and translation—into independent modules that run in series not only tends to accumulate errors across each stage, but also introduces noticeable latency and stuttering in actual use.

To address these pain points, SparkAudio1.0 completely breaks the boundaries of traditional processing. As a true speech foundation model, it is no longer limited to “speech-to-text”; instead, it has the ability to directly understand speech. It can “understand” multiple audio forms—such as human voice, environmental sounds, and music—at once, while simultaneously perceiving the deep semantics, emotional changes, and specific scenarios behind the sound.

Three Core Upgrades: Hearing More Broadly, Understanding More Deeply, and Independent Computing Power

Compared with previous speech technologies, the Spark speech foundation model SparkAudio1.0 has achieved comprehensive leaps across multiple key dimensions:

Hearing more: The model fully supports 99 languages and up to 202 dialects, and can efficiently handle extremely complex cross-language speech recognition and translation tasks.

Understanding more thoroughly: It leaps from basic speech recognition to multidimensional audio understanding, accurately grasping the deeper meaning of speech, thereby seamlessly supporting diverse industry application scenarios such as intelligent customer service, automotive cockpits, cross-language meetings, medical records, and content moderation.

Using more widely: The entire model was trained on a fully domestic computing infrastructure, providing developers and industries with a truly independent and controllable speech intelligence foundation.

At present, the related testing experience and API access channels for SparkAudio1.0 are fully open, marking a solid step forward for speech large models in commercialization and ecosystem openness.