Create README.md

README.md

@@ -0,0 +1,40 @@
+# Filipino-English Code-Switching Speech Dataset
+
+**Status: Under publication**
+
+A **Filipino-English (Taglish) read-speech dataset** created for training and evaluating automatic speech recognition (ASR) systems on code-switched speech. The corpus contains **3,555 manually validated utterances from 152 speakers**, totaling approximately **8.9 hours of audio**.
+
+|                  |                     |
+| ---------------- | ------------------- |
+| **Languages**    | Filipino + English  |
+| **Speech**       | Read, code-switched |
+| **Domain**       | Philippine news     |
+| **Utterances**   | 3,555               |
+| **Audio**        | FLAC, 16 kHz        |
+| **Train / Test** | 6.89 h / 2.04 h     |
+| **Split**        | Speaker-disjoint    |
+
+## Collection & Quality Control
+
+Reading prompts were drawn from Philippine news text and filtered specifically for **code-switching, readability, and appropriateness**. Recordings were collected through a web-based platform and reviewed by native Filipino-speaking annotators.
+
+Transcripts were manually normalized, including abbreviation expansion, number normalization, initialism handling, and correction of audio–text mismatches. Annotation disagreements were resolved through expert review. The dataset also provides **signal-to-noise ratio (SNR)** estimates for optional quality filtering.
+
+## ASR Results
+
+The dataset was evaluated using several pretrained ASR models.
+
+| Model                | Pretrained WER | After Fine-tuning |
+| -------------------- | -------------: | ----------------: |
+| **Whisper Large v3** |         13.62% |         **9.24%** |
+| MMS                  |         26.41% |        **17.60%** |
+| Qwen3-ASR            |         25.17% |        **20.13%** |
+| Gemma 4              |         16.18% |                 — |
+
+Fine-tuning on the dataset produced relative WER reductions of **32.16% for Whisper, 33.36% for MMS, and 20.02% for Qwen3-ASR**. In contrast, fine-tuning on general Filipino FLEURS data did not improve performance on the Taglish test set, highlighting the value of **domain-matched code-switched speech data**.
+
+## Intended Use & Limitations
+
+The dataset is intended for **Filipino-English ASR fine-tuning, benchmarking, and error analysis**.
+
+**The dataset and accompanying study are currently under publication. Final repository, license, and citation details will be provided upon release.**