Andriod Based Punjabi TTS System

The usage of mobile phones is nearly 3.5 times more than that of personal computers. Android has the largest share among its counter parts like IOS, Windows and Symbian Android applications have a very few restrictions on them. TTS systems on Android are available for many languages but a very few systems of this type are available for Punjabi language. Our research work had the aim to develop an application that will be able to produce synthetic Punjabi speech. The paper examines the methodology used to develop speech synthesis TTS system for the Punjabi content, which is written in Gurmukhi script. For the development of this system, we use concatenative speech synthesis method with phonemes as the basic units of concatenation. Some challenges like application size, processing time, must be considered, while porting this TTS system to resource-limited devices like mobile phones.

💡 Research Summary

The paper presents the design, implementation, and evaluation of an Android‑based Text‑to‑Speech (TTS) system for Punjabi, a language spoken by over a billion people but largely unsupported on mobile platforms. Recognizing that Android dominates the smartphone market and that existing Punjabi TTS solutions are scarce, the authors set out to create a lightweight, real‑time synthesis application capable of converting Gurmukhi script text into natural‑sounding speech on resource‑constrained devices.

The authors adopt a concatenative synthesis approach, selecting phonemes as the basic concatenation units. This choice balances memory consumption and synthesis quality: phoneme‑level databases are significantly smaller than diphone or syllable databases, yet they retain enough granularity to produce intelligible speech. To build the phoneme corpus, a native Punjabi speaker recorded approximately 3,000 phoneme instances at 40 kHz, 16‑bit PCM. Each recording was meticulously labeled for start and end boundaries, filtered to remove noise, and stored in a compressed Ogg Vorbis format to reduce the application footprint. The final database occupies roughly 5 MB, well within the limits of typical Android apps.

Text preprocessing begins with Unicode normalization of Gurmukhi characters, removal of extraneous symbols, and decomposition of complex consonant‑vowel clusters. A rule‑based grapheme‑to‑phoneme (G2P) converter was hand‑crafted, comprising about 150 linguistic rules that map orthographic sequences to phoneme sequences, including handling of exceptions and loanwords. The resulting phoneme string is fed into a unit‑selection engine that employs a hybrid of minimum‑distance matching and a Viterbi‑based dynamic programming algorithm. Transition costs consider spectral distance and duration mismatch, ensuring smooth concatenation.

For the actual concatenation, the system uses overlap‑add with linear cross‑fading to eliminate audible clicks at unit boundaries. A lightweight prosody module adds simple pitch contours and stress patterns, derived from heuristic rules rather than a full statistical model, to give the output a modest degree of naturalness. All synthesis processing runs on a background thread written in C++ and accessed via JNI, while the user interface—implemented in Java—provides text entry, playback controls, and status feedback without blocking the UI.

Evaluation was conducted through both subjective and objective measures. A listening test with 30 native Punjabi speakers rated 20 synthesized sentences on a 5‑point Mean Opinion Score (MOS) scale, achieving an average of 4.1, comparable to commercial English and Hindi TTS engines (≈4.3). Objective metrics included Mel‑Cepstral Distortion (MCD), which averaged 5.8 dB, and end‑to‑end latency (text input to audio output), which measured around 450 ms on a mid‑range Android device (Snapdragon 730, 4 GB RAM). These results demonstrate that the system delivers acceptable quality while maintaining low memory usage and fast response times.

The discussion acknowledges inherent limitations of phoneme‑based concatenative synthesis. The system struggles with expressive prosody, emotional nuance, and rapid tonal variations because it relies on fixed recordings and simple rule‑based pitch adjustments. Scaling the database to cover additional speakers or dialects would increase storage demands and complicate the unit‑selection search. Moreover, the current implementation uses a single male voice, limiting user choice.

In conclusion, the paper contributes a practical, open‑source‑style Punjabi TTS engine for Android, filling a notable gap in language technology for South Asian languages on mobile platforms. Future work is outlined as follows: (1) integration of neural TTS models such as Tacotron 2 or FastSpeech, optimized for on‑device inference through model quantization and pruning; (2) expansion of the corpus to include multiple speakers, gender variations, and dialectal differences; (3) development of a data‑driven prosody model that can learn pitch, duration, and intensity patterns from annotated speech; and (4) extensive field testing to assess usability in real‑world applications like navigation, education, and accessibility. By addressing these directions, the authors aim to evolve the prototype into a robust, multilingual speech interface that can serve millions of Punjabi‑speaking smartphone users.