Upload TRANSCRIPT_VARIANTS.md with huggingface_hub

TRANSCRIPT_VARIANTS.md

@@ -0,0 +1,45 @@
+# Transcript Variants
+
+This project contains two distinct sources of transcriptions for the same corpus of ~4885 sentences.
+
+## 1. Reference Transcript (PDF Based)
+- **Source**: Extracted from `atrinkti sak 2a.pdf` (provided by dataset authors).
+- **Processed File**: `atrinkti_sak_lines.txt`
+- **Characteristics**: 
+  - Represents the intended text.
+  - Contains accurate Unicode characters and stress marks (mapped to ASCII symbols `^` for Acute, `~` for Tilde, `` ` `` for Grave).
+  - Used as the source of truth for capitalization and spelling in the generated JSONs.
+
+## 2. Machine-Predicted Transcripts (Number Based)
+- **Origin**: These transcripts were generated by feeding raw (unstressed) text into a statistical accentuation model.
+  - The model is a deconstructed component of the **Liepa** synthesis engine.
+  - Source Code: [aleksas/phonology_engine](https://github.com/aleksas/phonology_engine)
+- **Differences**: 
+  - Since these are statistical predictions, the stress sometimes differs from the PDF reference.
+  - `edvardas` transcript is missing one sentence (ID 2435 in reference).
+
+## 3. Combined Output (`.json`)
+The generated JSON files (`aiste.json`, etc.) compile these sources.
+- **`id`**: The 1-based line number from the speaker's transcript file.
+- **`ref_id`**: The corresponding 1-based line number in the PDF reference (`atrinkti_sak_lines.txt`).
+  - Maps correctly even for `edvardas` (skips the missing line).
+  - Is `null` if the transcript line has no corresponding reference line (e.g. extra metadata lines at the end).
+- **`text`**: The sentence text with capitalization restored from the PDF reference (where available).
+- **`text_stressed_ascii_a`**: Text with **machine-predicted** stress, visualized using ASCII symbols (`^`, `~`, `` ` ``).
+- **`text_stressed_unicode_a`**: Text with **machine-predicted** stress, using standard Unicode combining diacritics.
+- **`text_stressed_ascii_b`**: Text with **PDF reference** stress, visualized using ASCII symbols.
+- **`text_stressed_unicode_b`**: Text with **PDF reference** stress, using standard Unicode combining diacritics.
+
+Note: Variants `_a` come from the statistical model (Transcript), while variants `_b` come from the original PDF source.
+
+
+## 4. Accent Mapping
+We process accents into two formats: ASCII-based (for readability/legacy) and Unicode combining diacritics.
+
+### Stress Types
+
+| Stress Name | ASCII Symbol | Unicode Combining | Example (ASCII) |
+|-------------|--------------|-------------------|-----------------|
+| **Grave**   | `` ` ``      | `\u0300`          | `a``            |
+| **Acute**   | `^`          | `\u0301`          | `a^`            |
+| **Tilde**   | `~`          | `\u0303`          | `a~`            |