Links table
Abstract and 1. Introduction
- Related work
- model
- Experiments
- Studying the eradication
- Conclusion, restrictions and risks
5. Studying the eradication
Semals and fast sound units make a series of S2ST claims for SechlessPressivelm. The natural question to be asked is the effectiveness of this immediate design for LM Training LM. So we have tried other debris strategies such as studying the eradication.
5.1 A series of ideas
This strategy mainly provides the same information as Cot Prompting for Model Training. Multi -task learning is used instead of multi -steps thinking. In Table 3, the “No Series of Thought” row shows a decrease of 9.82 and 4.38 in ASR-Bleu for ES-EN and HU-EN, respectively. It indicates that Cot helps the model while maintaining the indications better in the translation process.
To determine the importance of the semantic mentor in modeling, we do another strategic experience by removing the targeted semantic units from Cot Prompting. Specifically, the model is trained to predict directly with the targeted target audio units on the semantic source units and fast sound units. The row of “lack of a semantic mentor” in Table 3 shows a semantic deterioration with a decrease of 10.61 Asrbleu in ES-EN and 5.32 in HU-EN, indicating that the semantic claim plays an important role in providing semantic signals for S2ST modeling.
5.1.1 semantic router
The row of “lack of a semantic mentor” in Table 3 shows a decrease of 10.61 ASR-Beleu at ES-EN and 5.32 in HU-EN, indicating that the semantic claim plays an important role in providing semantic signals of S2ST modeling.
5.2 audio router
Moreover, a portion of the targeted speech is taken as an audio student in the semlessexpressivelm training, one of the important first action is the percentage. With Spanish take to English, for example, we train multiple models with three sets of claim ratio: (0.20, 0.25), (0.25, 0.3) and (0.30, 0.35). For each train sample, the claim ratio is chosen uniformly from the specified range. As for inference, we apply different guided ratios to test samples, and measure how to change ASR-Bleu and VSIM with it.
As shown in Figure 2, it achieves the range of training claim (0.25, 0.30) above ASR-Bleu with a test router of 0.3. Short audio claims cannot provide adequate audio information for translation, while the long audio claim may encourage the model to copy the claim and paste it as taken from the targeted speech. Trained models with (0.25, 0.30) and (0.30, 0.35) ASR-Bleu Achievement when the test claim ratio is set as 0.30. For the trained model with (0.20, 0.25), Asrbleu decreases when the test claim rate increases. We can see a steady improvement of VSIM with an increase in the test router in all three models.
