Intelligent Speech Interaction offers two customization tools to improve speech recognition accuracy when the default models fall short: hotwords for quick vocabulary boosts, and custom models for deeper, scenario-specific optimization.
To train and manage custom models in the console, see Manage custom models.
Choose a tool
| Hotwords | Custom models | |
|---|---|---|
| Best for | Specific words the model misrecognizes — names, locations, or industry-specific terms | Improving accuracy across an entire domain or scenario |
| When to use | The model fails on specific words in your business field | You need to recognize proper nouns and high-frequency words, or basic models cannot meet your business requirements |
| How it works | Add words to a hotword list to improve the effectiveness of recognizing specific words | Upload a training corpus to the self-learning platform; the model is retrained on your data |
| Learn more | Hotwords overview | Custom models overview |
Train a custom model for domain-specific recognition
The following example walks through a scenario where a team uses a custom model to improve recognition accuracy for domain-specific vocabulary.
Scenario: academic conference transcription
A redology research institute is hosting a seminar on *Dream of the Red Chamber* and wants to transcribe audio from guest speakers. Developers register an Alibaba Cloud account and activate Intelligent Speech Interaction. The team trains a custom model on the self-learning platform to handle terminology specific to the novel.
Select a basic model. The team selects the universal model as the starting point for customization.
Collect a training corpus. The team uses source text related to *Dream of the Red Chamber*. To prepare the training data, split the text at punctuation marks so that each sentence is stored as a separate line in the corpus file.
Train the custom model. On the self-learning platform, upload the training corpus and start the training job. After training completes, the model can accurately recognize domain-specific terms such as "Jia Baoyu" that the base model would otherwise miss.