端到端语音语言大模型（Lite）

What is End-to-End Speech Language Model Lite?

End-to-End Speech Language Model Lite is Baidu’s lower-cost, low-latency model for spoken dialogue. Instead of treating speech recognition, text generation and speech synthesis as completely separate user-facing steps, it uses Baidu’s Cross-Attention end-to-end speech-language architecture to process spoken interaction as a continuous conversation. The practical goal is a response that feels immediate when a person is speaking with an assistant, device or application.

The model is commercially available and belongs to Baidu’s End-to-End Speech Language Model family. The supplied documentation identifies two canonical API model identifiers: audio-mini-realtime-near for near-field scenarios and audio-mini-realtime-far for far-field scenarios. The model name is presented as “End-to-End Speech Language Model Lite,” while the API identifiers use Baidu’s audio-mini naming.

“Lite” should be understood as a positioning choice for real-time and cost-sensitive deployments, not as a claim that the model is a general-purpose replacement for every text or multimodal model. Its strongest fit is spoken interaction where response speed and audio handling matter more than broad reasoning depth.

Primary purpose and lineup positioning

Baidu positions this model for low-latency voice assistants, emotional-companionship applications, smart hardware, mobile applications and natural spoken dialogue. These use cases share a requirement: the system must listen, respond quickly and continue a conversational exchange without forcing the user through a rigid push-to-talk workflow.

The near-field and far-field identifiers indicate that Baidu provides variants for different audio-capture conditions. The supplied research does not publish a detailed performance table separating the two variants, so they should not be treated as different capability tiers without additional documentation. Their documented distinction is the intended acoustic setting: near-field use generally refers to audio captured close to the microphone, while far-field use refers to more distant capture.

This places Lite closer to a real-time conversational voice engine than to a conventional text-first large language model. A text-first model may be preferable for long documents, complex coding tasks or structured business workflows. Lite is more appropriate when the primary interface is speech and the user benefits from a prompt response.

Supported inputs and outputs

The model supports both audio and text input. Audio interaction uses PCM16 audio at 16 kHz, and Baidu documents real-time WebSocket interaction. The model can return both audio and text, including audio responses and audio transcripts. This allows an application to play a spoken reply while also retaining text for captions, logs, accessibility features or downstream display.

The model also supports interruption handling and server-side voice activity detection. Voice activity detection helps the service identify when a person has started or stopped speaking, while interruption handling is important for natural dialogue: a user can cut in rather than waiting silently for a long generated response to finish. These features are especially relevant for assistants embedded in devices or applications where conversational timing affects usability.

Latency and conversational behavior

Baidu lists approximately one-second end-to-end response latency. This is a provider-stated positioning claim rather than an independent benchmark, and actual latency will depend on network conditions, audio transmission, server load, application behavior and the time needed to generate a particular response. Even so, the claim clearly identifies the model’s design priority: rapid turn-taking.

Fast response is valuable in situations such as a voice-controlled appliance, a mobile assistant or a spoken customer interaction. A model that waits too long can make users repeat themselves or assume that the system has failed. Lite’s streaming interface, server-side voice activity detection and interruption support are designed around avoiding that kind of conversational friction.

The trade-off is that this model should not automatically be selected for tasks where maximum reasoning depth, extensive context or sophisticated code generation is the main requirement. The supplied research does not provide a context-window size or maximum output-token limit, so those values should be treated as unspecified rather than assumed to be unlimited.

Reasoning, coding and tool capabilities

The supplied editorial assessment gives End-to-End Speech Language Model Lite a reasoning score of 6 out of 10 and a coding score of 3 out of 10. These are editorial estimates based on the model’s documented real-time voice positioning, not Baidu-published benchmark results. They suggest that the model can handle ordinary conversational reasoning but is not primarily intended for difficult multi-step analysis or software development.

For example, Lite may be a sensible choice when a user asks a voice assistant a short factual question, requests a conversational response or interacts with a smart device. It is a weaker choice when the task involves generating a large codebase, debugging complex software, analyzing a long technical document or maintaining a structured multi-step workflow.

General tool or function calling is not documented as supported in the supplied research. The database marks tool use as unavailable. Baidu does list search enhancement as an optional, separately billed capability at ¥0.036 per call postpaid. That search feature should not be confused with broad agentic tool support: it is a specific optional enhancement rather than evidence that the model can freely call arbitrary application functions.

Pricing and cost structure

Baidu lists separate prices for audio and text tokens. Audio input costs ¥0.016 per 1,000 audio-input tokens, while text input costs ¥0.002 per 1,000 text-input tokens. Audio output costs ¥0.06 per 1,000 audio-output tokens, and text output costs ¥0.016 per 1,000 text-output tokens.

Baidu also advertises prepaid packages ranging from ¥70 for 2 million tokens to ¥25,000 for 1 billion tokens. The supplied information does not specify how each package allocates audio-input, text-input, audio-output and text-output usage, so package prices should not be converted into a single universal per-token rate.

The editorial cost score is 9 out of 10, reflecting the model’s low-cost positioning, especially for a real-time voice service. That score is subjective and should not be read as a provider guarantee. Actual spending depends on the mix of audio and text, the amount of generated audio, conversation length, interruptions and any separately billed search calls.

Main strengths and limitations

Strengths

• Real-time interaction: WebSocket streaming, server-side voice activity detection and interruption handling are designed for conversational turn-taking.
• Audio-native operation: The model accepts spoken input and produces spoken output rather than requiring an application to build the entire voice pipeline independently.
• Dual response modalities: Audio and text responses can support playback, captions, transcripts and application interfaces at the same time.
• Low-latency positioning: Baidu lists approximately one-second end-to-end response latency.
• Cost-sensitive deployment: Separate input and output prices, plus prepaid packages, support deployments where usage economics matter.
• Chinese speech focus: Baidu lists Mandarin and several regional Chinese dialects as supported speech varieties.

Limitations

• No published context or output ceiling in the supplied research: Context length and maximum output tokens are unspecified.
• Limited modality scope: Image and video understanding or generation are not documented, making the model unsuitable for visual applications.
• Weak coding fit: Its low editorial coding score reflects its speech-first purpose rather than a development-oriented design.
• No documented general tool calling: Applications needing function execution or broad agent workflows should use an option with explicit tool support.
• Not a batch-processing model: The supplied positioning emphasizes live WebSocket sessions, not batch API workloads.
• Audio output can cost more: Audio-output tokens are priced higher than audio-input tokens, so applications that generate long spoken responses should monitor usage carefully.

When to choose End-to-End Speech Language Model Lite

Choose this model when speech is the central interface and fast turn-taking matters. It is a strong candidate for a Mandarin-speaking voice assistant, an interactive smart appliance, a mobile voice feature, an emotional-companionship experience or another application where users expect to interrupt and resume a spoken conversation naturally.

It is also a reasonable choice when an application needs both audio playback and text transcripts. For example, a device can play the generated PCM16 response while displaying the transcript on screen, or a mobile application can retain the text for accessibility and conversation history.

Consider another type of model when the main task is long-form text generation, image or video understanding, advanced coding, complex tool orchestration, fine-tuning or batch processing. A text-first reasoning model is likely to be a better fit for deep analysis, while a vision-language model is more appropriate for images or video. The supplied research does not identify a specific Baidu sibling model that should replace Lite for these tasks, so those comparisons are best understood as capability-based rather than direct model-to-model recommendations.

Bottom line

End-to-End Speech Language Model Lite is a focused Baidu product for responsive spoken dialogue. Its value comes from the combination of audio input and output, real-time WebSocket streaming, interruption handling, voice activity detection and a low-latency commercial positioning. It is not presented as a general-purpose model for every modality or workload. The clearest reason to choose it is the need for affordable, natural and fast voice interaction; the clearest reason to choose something else is a requirement for visual understanding, extensive reasoning, coding, general tool use or documented large-context processing.