Standardizing Numbers and Proper Nouns: Handling Entities in Vietnamese Speech-to-Text
Learn how to standardize numbers and proper nouns in Vietnamese STT. Master post-processing, NER, and code-switching for accurate, readable text.
In this article
In speech-to-text (STT) systems, converting audio into raw text is only the first step. The real challenge for developers and users lies in handling STT entities to ensure the output is meaningful, readable, and accurate. For Vietnamese, issues with numbers, currency, and proper nouns often lead to misinterpretations without a robust post-processing speech to text workflow. This article explores practical techniques to normalize data, turning raw character strings into valuable information.
Why Standardize Numbers and Proper Nouns in STT?
Raw output from speech recognition models is typically a sequence of unanalyzed words. For example, the spoken phrase "I have three million dong" might be transcribed correctly or misheard due to similar phonetics.
For Vietnamese number standardization, the distinction between "twelve" (12) and "twelve o'clock" (12:00) is critical. Without context handling, systems cannot differentiate between units of measurement, currency, or time. Similarly, proper nouns like "Hai Phong," "Thu Thiem," or international brand names spoken with a Vietnamese accent are prone to spelling errors or missing diacritics.
Key Post-Processing Techniques
To improve post-recognition text quality, apply Natural Language Processing (NLP) steps in your post-processing pipeline.
1. Context and Entity Analysis (NER)
Named Entity Recognition (NER) is crucial. The system must identify specific entity groups:
- Quantities and Units: Recognize numbers accompanied by units (dong, thousand, million, km, kg, etc.).
- Time: Identify dates, months, years, hours, and minutes.
- Proper Nouns: Detect person names, locations, and organization names.
When a numeric entity is detected, the system references a Vietnamese number dictionary to convert it to Arabic numerals or keep it in word form, depending on your configuration. For instance, "two hundred fifty thousand" is normalized to "250,000."
2. Proper Noun and Spelling Handling
Vietnamese is a diacritic-rich language, but in speech, tones are not always distinct enough for perfect transcription. While STT models are trained to predict diacritics based on context, proper nouns remain a challenging area for accuracy.
Practical advice:
- Build a Custom Vocabulary for product names, employee names, and specific locations in your industry.
- Use algorithms like Levenshtein distance to compare recognition results against your dictionary, suggesting the most accurate correction.
3. Output Formatting
After entity recognition, the final step is reformatting the sentence.
- Capitalization: Automatically capitalize the first letter of sentences and words after periods.
- Number Formatting: Apply thousand-separator rules (e.g., 1,000,000 instead of 1000000) for large numbers.
- Code-Switching: For sentences mixing Vietnamese and English (e.g., "send an email to Mr. Nam"), the system should preserve the English structure while normalizing the Vietnamese parts.
Real-World Processing Example
Consider this snippet from a business meeting: "Total Q3 revenue was five hundred million dong, growing twenty percent compared to the previous quarter."
Step 1: Raw STT Output
tong doanh thu quy ba la nam tram trieu dong, tang truong hai muoi phan tram so voi quy truoc
Step 2: Entity Processing and Normalization
- Detect number: "nam tram trieu" -> 500,000,000
- Detect unit: "dong" -> VND
- Detect percentage: "hai muoi phan tram" -> 20%
- Normalize time/proper noun: "quy ba" -> Q3
Step 3: Final Output
This process clearly demonstrates why post-processing speech to text is as important as the underlying recognition model quality.
Optimizing with AIVISION Speech-to-Text
At AIVISION, we understand that the challenge isn't just hearing correctly, but understanding Vietnamese context. Our Speech-to-Text model is designed specifically for Vietnamese, handling code-switching (mixed Vietnamese and English) effectively and providing word-level timestamps. This allows downstream NLP systems to easily perform STT entity handling.
Trained on a corpus of over 690,000 hours of Vietnamese speech, AIVISION's model achieves high accuracy across standard test sets, including specialized fields like healthcare and business. This creates a solid foundation for subsequent number and proper noun standardization steps.
If you are building a virtual assistant, automated meeting minutes, or voice chatbot, integrating a robust STT pipeline with precise post-processing is the key to success. You can explore more technology solutions in our Blog.
Tips for Implementation
- Don't Skip Audio Cleaning: Good noise reduction and Voice Activity Detection (VAD) help the STT model focus on speech, reducing background interference.
- Customize Your Dictionary: Always provide a list of industry-specific terms to improve proper noun accuracy.
- Test Multiple Contexts: Evaluate Vietnamese number standardization performance across different scenarios (transactions, healthcare, engineering) to adjust parameters effectively.
Conclusion
Entity handling and text normalization are essential steps to transform raw audio data into useful information. By combining a high-quality STT model with NLP techniques like NER and formatting, you can significantly enhance the user experience.
AIVISION is committed to providing optimized Speech-to-Text solutions for Vietnamese, supporting both real-time streaming and file transcription via REST API. Start today with a free trial—no credit card required—to experience accurate recognition and context handling.
Visit Start free to create an account and get started, or Contact us to discuss a solution tailored to your business needs.
Frequently asked questions
Why does Vietnamese STT often make errors with numbers?
Vietnamese has various ways of reading numbers and complex measurement units. If the model lacks sufficient training data for numbers or lacks a post-processing step to identify context, the system easily confuses numbers with similar sounds.
How can I improve accuracy for proper nouns in STT?
You should use the Custom Vocabulary feature if supported by the provider, or build a post-processing layer to compare and correct words that are close to your fixed list of proper nouns.
Does AIVISION support mixed Vietnamese and English in the same sentence?
Yes, AIVISION Speech-to-Text handles code-switching (mixing Vietnamese and English) effectively, ensuring natural and accurate output for international or technical conversations.
Try AIVISION's Vietnamese speech AI
Speech-to-Text, Text-to-Speech and LLM, free.
Start free → Contact