leeoxiang/qwen3-0.6b-zh-tn

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 12, 2026License:cc-by-nc-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The leeoxiang/qwen3-0.6b-zh-tn model is a 0.8 billion parameter Qwen3-Base fine-tune specifically designed for Chinese Text Normalization (TN). It converts 14 categories of non-standard Chinese text (NSW) into their spoken forms for Text-to-Speech (TTS) systems. This model outputs an edit script rather than rewriting full sentences, enabling low-latency processing and minimizing hallucinations by only modifying specific spans.

Loading preview...

Overview

The leeoxiang/qwen3-0.6b-zh-tn is a specialized 0.8 billion parameter model, fine-tuned from Qwen3-0.6B-Base, for Chinese Text Normalization (TN). Its primary function is to convert non-standard Chinese words (NSW) such as numbers, dates, times, currency, and percentages into their appropriate spoken forms for Text-to-Speech (TTS) applications. Unlike models that rewrite entire sentences, this model generates an "edit script" indicating which text segments to modify and how, which is then applied by a parser. This approach ensures low latency and reduces the risk of hallucinations.

Key Capabilities

  • Span-Edit Text Normalization: Focuses on editing specific non-standard segments rather than full sentence generation.
  • Comprehensive NSW Coverage: Handles 14 distinct categories of non-standard Chinese text, including numerical expressions, dates, times, and financial figures.
  • High Accuracy: Achieves 99.79% sentence accuracy on synthetic blind tests and 96.5% on real-world encyclopedia corpus, with acceptable rates even higher.
  • Low Latency: Designed for efficiency, with P50 latency of 20.6ms and P99 of 63ms on H200 (vLLM, FP8, bs=1).
  • Customizable Output: Provides an edit script format (anchor->reading) that allows for flexible integration into TTS frontends.

Good for

  • TTS Pre-processing: Ideal for preparing Chinese text for speech synthesis, ensuring natural and correct pronunciation of non-standard elements.
  • Real-time Applications: Its low-latency design makes it suitable for applications requiring quick text normalization.
  • Specific Chinese TN Tasks: Excels in converting numerical and other non-standard expressions into their spoken equivalents, adhering to specific reading conventions (e.g., '1' as 'yao' in phone numbers).

Limitations

  • Single Task Focus: Exclusively performs text normalization; it does not handle polyphonic disambiguation or prosody.
  • Sentence Length: Best suited for sentence-level inputs (up to 150 characters); longer texts should be segmented.
  • Licensing: Trained partly on non-commercial data, the model is released under CC-BY-NC-4.0, requiring commercial users to replace or license the relevant data.

Finetunes
1 models