Model Introduction

Doubao Speech Recognition Large Model 2.0 is a newly upgraded model based on large-model speech recognition capabilities, leveraging industry-leading self-developed speech recognition technology and massive speech industry data advantages. With more sensitive ears and a smarter brain, recognition accuracy is further improved. It supports real-time recognition of long audio into text, achieving a "speak and see text simultaneously" effect, suitable for real-time speech recognition scenarios such as real-time meeting subtitles, live streaming subtitles, intelligent outbound calls, and more.

Product Advantages

• Ultra-high accuracy: Compared with traditional models, recognition error rate is reduced by 30%, and in vertical domains such as music, technology, education, and healthcare, the error rate is reduced by more than 50%.
• Improved recognition in complex scenarios: Supports multilingual and multi-dialect speech recognition, with accent error rate reduced by 60% and noise/background voice reduced by 30%–50%.
• More human-like interactive experience: The large model can leverage context, user input, and background information to enable the "ears" to think, delivering recognition results that better fit the context.

Feature Specifications

Feature	Doubao Streaming Speech Recognition Model
Sensitive word filtering	✅
Intelligent sentence segmentation	✅
Character/word timestamp	✅
Punctuation prediction	✅
Semantic smoothing (Chinese, English)	✅
Number normalization (ITN)	✅
VAD-based sentence segmentation	✅
Automatic speaker diarization (Chinese & English)	✅
Context (text, image)	✅ (2.0 supports image)
Forced end-of-speech timeout	✅
Output speech pauses, sentence/word segmentation info	✅
Gender detection	✅
Sentence info with speech rate	✅
Sentence info with volume	✅
Language detection	✅
Emotion detection	✅
Hot word correction - platform level	✅
Hot word correction - request level	✅
Regex replacement words	✅
First character return acceleration	✅
First character return acceleration rate	✅
Input audio format	Supports pcm, opus, mp3 formats
Sample rate	No requirement
Music function call	Bidirectional streaming optimized version - supported with second pass enabled
POI function call	Bidirectional streaming optimized version - supported with second pass enabled
Supported languages	Chinese & English (second pass supports Chinese, English, and dialects)

Billing

• Billing method: Billed by duration
• Billing unit: Hour
• Pre-freeze basis: Quota is pre-frozen based on the requested audio duration before the call
• Settlement basis: Settled based on the actual audio duration returned by the upstream

Note

• Only successfully recognized requests are billed.

Application Scenarios

• Voice Interaction: Provides a voice input channel for human-computer interaction, converting speech to text in real time as input, enabling fast and convenient interaction with devices/hardware/applications.
• Content Review & Quality Inspection: Converts recordings to text, analyzes text through quality inspection rules to promptly detect and intervene with non-compliant content; or monitors and analyzes content to discover potential business opportunities.
• Meeting & Interview Transcription: Recognizes meeting and interview audio in real time or asynchronously into text, automatically segments voiced portions for recognition, reducing costs and improving efficiency; simultaneously auto-segments to effectively improve meeting content recording efficiency.
• Game Voice Input: For game voice input and mobile input method scenarios, supports a "speak and see text simultaneously" effect, greatly reducing user effort in text communication and improving communication efficiency.
• Classroom Content Analysis: Recognizes classroom recordings, restores classroom scenes through text, analyzes teaching content, and improves teaching quality.
• Audio/Video Subtitles: Supports automatic conversion of speech and lyrics in audio/video into text, generating subtitle content corresponding to the audio/video with one click. Suitable for video editing, video watching, video conferencing, and other scenarios.