Commit
·
9103cb2
1
Parent(s):
8814ae3
correct the readme
Browse files
README.md
CHANGED
|
@@ -1,4 +1,6 @@
|
|
| 1 |
---
|
|
|
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
base_model: Qwen/Qwen2.5-7B
|
| 4 |
library_name: peft
|
|
@@ -8,8 +10,12 @@ tags:
|
|
| 8 |
- french
|
| 9 |
- qwen2.5
|
| 10 |
- lora
|
|
|
|
|
|
|
| 11 |
---
|
| 12 |
-
|
|
|
|
|
|
|
| 13 |
|
| 14 |
This is the second-stage LoRA adapter for **French SSML generation**, converting *pause-annotated text* into full SSML markup with `<break>` tags.
|
| 15 |
|
|
@@ -52,6 +58,9 @@ Output: Bonjour<break time="250ms"/> comment vas-tu ?
|
|
| 52 |
|
| 53 |
---
|
| 54 |
|
|
|
|
|
|
|
|
|
|
| 55 |
```python
|
| 56 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 57 |
from peft import PeftModel
|
|
@@ -69,8 +78,10 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))```
|
|
| 69 |
|
| 70 |
---
|
| 71 |
|
|
|
|
| 72 |
## 🧪 Evaluation Summary
|
| 73 |
|
|
|
|
| 74 |
| Metric | Value |
|
| 75 |
|--------------------------|---------------|
|
| 76 |
| Pause Insertion Accuracy | 87.3% |
|
|
@@ -79,9 +90,12 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))```
|
|
| 79 |
|
| 80 |
Evaluation was performed on a held-out French validation set with annotated SSML pauses. Mean Opinion Score (MOS) improvements were assessed using TTS outputs rendered with Azure Henri voice and rated by 30 native French speakers.
|
| 81 |
|
|
|
|
| 82 |
---
|
| 83 |
|
| 84 |
-
|
|
|
|
|
|
|
| 85 |
|
| 86 |
This LoRA adapter was trained on a corpus of ~4,500 French utterances. Input texts were annotated with symbolic pause indicators (e.g., `#250` for 250ms), automatically aligned using a combination of Whisper-Kyutai timestamping and F0/syntactic heuristics.
|
| 87 |
|
|
@@ -111,7 +125,7 @@ Training was performed using the [Unsloth](https://github.com/unslothai/unsloth)
|
|
| 111 |
|
| 112 |
---
|
| 113 |
|
| 114 |
-
## ⚠️
|
| 115 |
|
| 116 |
- Only `<break>` tags are supported; no pitch, rate, or emphasis control yet.
|
| 117 |
- Pause accuracy is sensitive to punctuation and malformed inputs.
|
|
@@ -120,6 +134,8 @@ Training was performed using the [Unsloth](https://github.com/unslothai/unsloth)
|
|
| 120 |
🔗 [`nassimaODL/ssml-text2breaks-fr-lora`](https://huggingface.co/nassimaODL/ssml-text2breaks-fr-lora)
|
| 121 |
|
| 122 |
---
|
|
|
|
|
|
|
| 123 |
@inproceedings{ould-ouali2025improving,
|
| 124 |
author = {Nassima Ould-Ouali and Awais Sani and Tim Luka Horstmann and Jonah Dauvet and Ruben Bueno and Éric Moulines},
|
| 125 |
title = {Improving French Synthetic Speech Quality via SSML Prosody Control},
|
|
|
|
| 1 |
---
|
| 2 |
+
|
| 3 |
+
|
| 4 |
license: apache-2.0
|
| 5 |
base_model: Qwen/Qwen2.5-7B
|
| 6 |
library_name: peft
|
|
|
|
| 10 |
- french
|
| 11 |
- qwen2.5
|
| 12 |
- lora
|
| 13 |
+
|
| 14 |
+
|
| 15 |
---
|
| 16 |
+
|
| 17 |
+
# 🗣️ ssml-break2ssml-fr-lora
|
| 18 |
+
|
| 19 |
|
| 20 |
This is the second-stage LoRA adapter for **French SSML generation**, converting *pause-annotated text* into full SSML markup with `<break>` tags.
|
| 21 |
|
|
|
|
| 58 |
|
| 59 |
---
|
| 60 |
|
| 61 |
+
### How to run the code
|
| 62 |
+
|
| 63 |
+
|
| 64 |
```python
|
| 65 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 66 |
from peft import PeftModel
|
|
|
|
| 78 |
|
| 79 |
---
|
| 80 |
|
| 81 |
+
|
| 82 |
## 🧪 Evaluation Summary
|
| 83 |
|
| 84 |
+
|
| 85 |
| Metric | Value |
|
| 86 |
|--------------------------|---------------|
|
| 87 |
| Pause Insertion Accuracy | 87.3% |
|
|
|
|
| 90 |
|
| 91 |
Evaluation was performed on a held-out French validation set with annotated SSML pauses. Mean Opinion Score (MOS) improvements were assessed using TTS outputs rendered with Azure Henri voice and rated by 30 native French speakers.
|
| 92 |
|
| 93 |
+
|
| 94 |
---
|
| 95 |
|
| 96 |
+
|
| 97 |
+
## 📚 Training Data
|
| 98 |
+
|
| 99 |
|
| 100 |
This LoRA adapter was trained on a corpus of ~4,500 French utterances. Input texts were annotated with symbolic pause indicators (e.g., `#250` for 250ms), automatically aligned using a combination of Whisper-Kyutai timestamping and F0/syntactic heuristics.
|
| 101 |
|
|
|
|
| 125 |
|
| 126 |
---
|
| 127 |
|
| 128 |
+
## ⚠️ Limitations
|
| 129 |
|
| 130 |
- Only `<break>` tags are supported; no pitch, rate, or emphasis control yet.
|
| 131 |
- Pause accuracy is sensitive to punctuation and malformed inputs.
|
|
|
|
| 134 |
🔗 [`nassimaODL/ssml-text2breaks-fr-lora`](https://huggingface.co/nassimaODL/ssml-text2breaks-fr-lora)
|
| 135 |
|
| 136 |
---
|
| 137 |
+
|
| 138 |
+
## 📖 Citation
|
| 139 |
@inproceedings{ould-ouali2025improving,
|
| 140 |
author = {Nassima Ould-Ouali and Awais Sani and Tim Luka Horstmann and Jonah Dauvet and Ruben Bueno and Éric Moulines},
|
| 141 |
title = {Improving French Synthetic Speech Quality via SSML Prosody Control},
|