docs/Indian_sounds_difference.md

Comprehensive Guide to Indian English Phonetics for ASR Systems

This document outlines the articulatory phonetic rules that govern Indian English pronunciation, specifically designed for Automatic Speech Recognition (ASR) pipelines and Goodness of Pronunciation (GoP) assessment tools. It maps standard Western acoustic expectations to the physical realities of Indian English speech.

---

1. Core Architectural Terminology

To understand how phonetic mapping impacts a machine learning pipeline, these core concepts must be defined:

• ASR (Automatic Speech Recognition): The overarching AI pipeline that converts spoken audio waveforms into written text.
• Acoustic Model (Wav2Vec2): The "ear" of the system. It does not understand grammar; it processes raw audio waves and predicts microscopic speech sounds (phonemes) millisecond by millisecond.
• Lexicon / Dictionary: The translation rulebook. It maps the phonetic sounds outputted by the Acoustic Model back to written English words (e.g., mapping t̪ ɒ ʈ to the word "thought").
• Classification Head: The final mathematical layer of the neural network that outputs a probability score for every possible sound token in the vocabulary.

---

2. Phonetic Alphabets & Token Sets

Speech datasets utilize different alphabetic frameworks to represent sound. Selecting the right framework determines the Acoustic Model's sensitivity to pronunciation errors.

Framework	Origin	Description	Example (Word: That)
ARPAbet	CMUdict (American)	Standard US English representation using uppercase ASCII letters.	DH AE T
Detailed IPA	Scientific Standard	High-definition Unicode symbols tracking exact tongue and lip placement.	d̪ a ʈ
IE-CPS	CDAC (Indian)	Common Phone Set using ASCII characters to safely represent Indian sounds.	dx a t

Strategic Note: Detailed IPA provides the high-definition acoustic granularity required for precise pronunciation grading, while IE-CPS is optimized for broad, cross-language text generation and backend system safety.

---

3. The Indian English Phonological Rules

Indian English is heavily influenced by the articulatory frameworks of indigenous Indo-Aryan and Dravidian languages. When Indian speakers speak English, they naturally shift Western sounds to fit this native consonant grid.

Rule 1: Dentalization (The "TH" Shift)

Linguistic Shift: Interdental Fricative $\rightarrow$ Dental Stop

• The Western Baseline: In American/British English, the "th" sounds (as in thought or that) are fricatives. The tongue tip rests lightly between the front teeth, and air is pushed through the gap to create a continuous hiss.
• The Indian Shift: Indian English speakers articulate this as a Dental Stop. The tongue completely blocks the airflow by pressing flat against the back of the upper teeth, followed by a sudden release or "pop" of air (similar to the Hindi त or द).
• Acoustic Impact: The audio waveform shifts from a long block of high-frequency friction noise to a sudden spike in energy.
• Token Mapping:
  • Standard ARPAbet: TH / DH
  • Detailed IPA: t̪ / d̪ (The subscript bridge explicitly indicates tongue-on-teeth contact)
  • IE-CPS: tx / dx

Rule 2: Retroflexion (The "T/D" Shift)

Linguistic Shift: Alveolar Stop $\rightarrow$ Retroflex Stop

• The Western Baseline: Standard T and D sounds (as in top or dog) are alveolar stops. The tongue taps the bumpy ridge directly behind the upper front teeth.
• The Indian Shift: Because the dental position is occupied by the shifted "th" sounds, the standard T and D are pushed backward into the retroflex position. The tip of the tongue curls backward to strike the hard roof of the mouth (similar to the Hindi ट or ड).
• Acoustic Impact: The curled tongue creates a larger hollow cavity in the mouth, which visually drops the frequency of the third formant ($F_3$) on a spectrogram.
• Token Mapping:
  • Standard ARPAbet: T / D
  • Detailed IPA: ʈ / ɖ (The rightward downward hook represents the physical curling of the tongue)
  • IE-CPS: t / d

Rule 3: Monophthongization (The Vowel Shift)

Linguistic Shift: Gliding Diphthongs $\rightarrow$ Long Monophthongs

• The Western Baseline: Words like cake (/keɪk/) or boat (/boʊt/) utilize diphthongs. The jaw and tongue start at one vowel position and glide to another mid-syllable, changing the shape of the sound.
• The Indian Shift: Indian English utilizes pure, sustained monophthongs. The tongue and jaw remain completely stationary, holding a single flat vowel position for a longer duration (e.g., cake becomes /keːk/).
• Acoustic Impact: Instead of slanted, diagonal frequency bands on a spectrogram (indicating a glide), the audio features perfectly flat, horizontal frequency bars that endure over time.
• Token Mapping:
  • Standard ARPAbet: EY / OW
  • Detailed IPA: eː / oː (The triangular colon is the chroneme, indicating extended acoustic duration)
  • IE-CPS: ey / oh

---

4. Conclusion for AI Architecture

A standard ASR model trained on Western phonetic rules will consistently generate false errors for Indian speakers. It will penalize a perfectly pronounced Indian that (/d̪ a ʈ/) because the model's loss function mathematically expects a Western fricative and alveolar stop.

By updating the Grapheme-to-Phoneme (G2P) dictionaries to target the Detailed IPA framework, the Acoustic Model learns to accept Dentalization, Retroflexion, and Monophthongization as the correct ground-truth targets. This ensures the Goodness of Pronunciation (GoP) algorithms accurately grade speech clarity rather than penalizing regional accents.