Update README.md
Browse files
README.md
CHANGED
|
@@ -23,7 +23,6 @@ pipeline_tag: automatic-speech-recognition
|
|
| 23 |
**Base Model:** `openai/whisper-small` (fine-tuned for Swahili)
|
| 24 |
|
| 25 |
---
|
| 26 |
-
|
| 27 |
## 🌍 Overview
|
| 28 |
|
| 29 |
**SALAMA-STT** (Speech-to-Text) is the **first module** of the **SALAMA Framework** — a modular end-to-end **speech-to-speech AI system** built for African languages.
|
|
@@ -50,7 +49,6 @@ The model was fine-tuned on the **Mozilla Common Voice 17.0 Swahili** dataset, e
|
|
| 50 |
| Languages | Swahili (`sw`), English (`en`) |
|
| 51 |
|
| 52 |
---
|
| 53 |
-
|
| 54 |
## 📚 Dataset
|
| 55 |
|
| 56 |
| Dataset | Description | Purpose |
|
|
@@ -60,7 +58,6 @@ The model was fine-tuned on the **Mozilla Common Voice 17.0 Swahili** dataset, e
|
|
| 60 |
| Common Voice validation split | 2.3 hours | Evaluation |
|
| 61 |
|
| 62 |
---
|
| 63 |
-
|
| 64 |
## 🧠 Model Capabilities
|
| 65 |
|
| 66 |
- Speech-to-text transcription in **Swahili**
|
|
@@ -70,7 +67,6 @@ The model was fine-tuned on the **Mozilla Common Voice 17.0 Swahili** dataset, e
|
|
| 70 |
- Provides timestamped segment transcriptions
|
| 71 |
|
| 72 |
---
|
| 73 |
-
|
| 74 |
## 📊 Evaluation Metrics
|
| 75 |
|
| 76 |
| Metric | Baseline (Whisper-small) | Fine-tuned (SALAMA-STT) | Improvement |
|
|
@@ -82,7 +78,6 @@ The model was fine-tuned on the **Mozilla Common Voice 17.0 Swahili** dataset, e
|
|
| 82 |
> Evaluation conducted on a 2-hour held-out Swahili validation set from Common Voice.
|
| 83 |
|
| 84 |
---
|
| 85 |
-
|
| 86 |
## ⚙️ Usage (Python Example)
|
| 87 |
|
| 88 |
Below is a quick example for Swahili speech transcription using this model:
|
|
@@ -113,7 +108,6 @@ print(result["text"])
|
|
| 113 |
> *“Karibu kwenye mfumo wa SALAMA unaosaidia kutambua na kuelewa sauti ya Kiswahili kwa usahihi mkubwa.”*
|
| 114 |
|
| 115 |
---
|
| 116 |
-
|
| 117 |
## 🔍 Model Performance Summary
|
| 118 |
|
| 119 |
| Dataset | Metric | Score |
|
|
@@ -123,7 +117,6 @@ print(result["text"])
|
|
| 123 |
| Local Swahili Test Set | Accuracy | **95.4%** |
|
| 124 |
|
| 125 |
---
|
| 126 |
-
|
| 127 |
## ⚡ Key Features
|
| 128 |
|
| 129 |
- 🎙️ **Accurate Swahili ASR** trained on diverse voices
|
|
@@ -133,7 +126,6 @@ print(result["text"])
|
|
| 133 |
- 🚀 **Fast inference optimized with FP16 precision**
|
| 134 |
|
| 135 |
---
|
| 136 |
-
|
| 137 |
## 🚫 Limitations
|
| 138 |
|
| 139 |
- May misinterpret **code-mixed (Swahili-English)** speech
|
|
@@ -142,11 +134,9 @@ print(result["text"])
|
|
| 142 |
- Performance may decline on **non-native Swahili speakers**
|
| 143 |
|
| 144 |
---
|
| 145 |
-
|
| 146 |
## 🔗 Related Models
|
| 147 |
|
| 148 |
| Model | Description |
|
| 149 |
|--------|-------------|
|
| 150 |
| [`EYEDOL/salama-llm`](https://huggingface.co/EYEDOL/salama-llm) | Swahili instruction-tuned LLM for reasoning and dialogue |
|
| 151 |
| [`EYEDOL/salama-tts`](https://huggingface.co/EYEDOL/salama-tts) | Swahili text-to-speech (VITS) model for natural speech synthesis |
|
| 152 |
-
|
|
|
|
| 23 |
**Base Model:** `openai/whisper-small` (fine-tuned for Swahili)
|
| 24 |
|
| 25 |
---
|
|
|
|
| 26 |
## 🌍 Overview
|
| 27 |
|
| 28 |
**SALAMA-STT** (Speech-to-Text) is the **first module** of the **SALAMA Framework** — a modular end-to-end **speech-to-speech AI system** built for African languages.
|
|
|
|
| 49 |
| Languages | Swahili (`sw`), English (`en`) |
|
| 50 |
|
| 51 |
---
|
|
|
|
| 52 |
## 📚 Dataset
|
| 53 |
|
| 54 |
| Dataset | Description | Purpose |
|
|
|
|
| 58 |
| Common Voice validation split | 2.3 hours | Evaluation |
|
| 59 |
|
| 60 |
---
|
|
|
|
| 61 |
## 🧠 Model Capabilities
|
| 62 |
|
| 63 |
- Speech-to-text transcription in **Swahili**
|
|
|
|
| 67 |
- Provides timestamped segment transcriptions
|
| 68 |
|
| 69 |
---
|
|
|
|
| 70 |
## 📊 Evaluation Metrics
|
| 71 |
|
| 72 |
| Metric | Baseline (Whisper-small) | Fine-tuned (SALAMA-STT) | Improvement |
|
|
|
|
| 78 |
> Evaluation conducted on a 2-hour held-out Swahili validation set from Common Voice.
|
| 79 |
|
| 80 |
---
|
|
|
|
| 81 |
## ⚙️ Usage (Python Example)
|
| 82 |
|
| 83 |
Below is a quick example for Swahili speech transcription using this model:
|
|
|
|
| 108 |
> *“Karibu kwenye mfumo wa SALAMA unaosaidia kutambua na kuelewa sauti ya Kiswahili kwa usahihi mkubwa.”*
|
| 109 |
|
| 110 |
---
|
|
|
|
| 111 |
## 🔍 Model Performance Summary
|
| 112 |
|
| 113 |
| Dataset | Metric | Score |
|
|
|
|
| 117 |
| Local Swahili Test Set | Accuracy | **95.4%** |
|
| 118 |
|
| 119 |
---
|
|
|
|
| 120 |
## ⚡ Key Features
|
| 121 |
|
| 122 |
- 🎙️ **Accurate Swahili ASR** trained on diverse voices
|
|
|
|
| 126 |
- 🚀 **Fast inference optimized with FP16 precision**
|
| 127 |
|
| 128 |
---
|
|
|
|
| 129 |
## 🚫 Limitations
|
| 130 |
|
| 131 |
- May misinterpret **code-mixed (Swahili-English)** speech
|
|
|
|
| 134 |
- Performance may decline on **non-native Swahili speakers**
|
| 135 |
|
| 136 |
---
|
|
|
|
| 137 |
## 🔗 Related Models
|
| 138 |
|
| 139 |
| Model | Description |
|
| 140 |
|--------|-------------|
|
| 141 |
| [`EYEDOL/salama-llm`](https://huggingface.co/EYEDOL/salama-llm) | Swahili instruction-tuned LLM for reasoning and dialogue |
|
| 142 |
| [`EYEDOL/salama-tts`](https://huggingface.co/EYEDOL/salama-tts) | Swahili text-to-speech (VITS) model for natural speech synthesis |
|
|
|