Praxy Voice: Voice-Prompt Recovery + BUPS for Commercial-Class Indic TTS from a Frozen Non-Indic Base at Zero Commercial-Training-Data Cost

Venkata Pushpak Teja Menta

Source-linked AI summary

Praxy Voice: Voice-Prompt Recovery + BUPS for Commercial-Class Indic TTS from a Frozen Non-Indic Base at Zero Commercial-Training-Data Cost

Venkata Pushpak Teja Menta

arXiv:2604.25441v1Apr 28, 2026cs.SDcs.CLeess.AS

TL;DR

Indic TTS typically requires costly native-model training or dependence on closed commercial systems. Praxy Voice uses BUPS, text-head LoRA adaptation, and voice-prompt recovery to reach commercial-class output, with 26.7% Telugu retroflex collapse and 0.025 Hindi LLM-WER.

Problem
Indic production TTS has required either costly native-model training or per-call dependence on closed commercial systems.
Method
Praxy Voice combines deterministic BUPS romanisation, text-head LoRA adaptation, and shared voice-prompt recovery with three-branch routing.
Results
26.7% Telugu retroflex collapse, 71% Tamil-zha collapse, and 0.025 Hindi LLM-WER place Praxy in the commercial performance pack; code-mix LLM-WER reaches 0.14–0.27.
Takeaways & Limitations
A frozen non-Indic-native base can reach commercial-class output on three Indic languages through minimum intervention, while Hindi requires vanilla rather than LoRA routing.
Takeaways & Limitations
The reported comparisons come from 10-utterance pilots whose small sample sizes cannot statistically separate 5-percentage-point differences.

Abstract

from arXiv · show

Commercial TTS systems produce near-native Indic audio, but the best open-source bases (Chatterbox, Indic Parler-TTS, IndicF5) trail them on measured phonological dimensions, and the most widely adopted multilingual base (Chatterbox, 23 languages) does not even tokenise Telugu or Tamil. We ask: what is the minimum intervention that brings such a non-Indic-native base to commercial-class output on Telugu, Tamil, and Hindi, without training a new acoustic decoder and without any commercial TTS training data? We combine three pieces: (1) BUPS, a Brahmic Unified Phoneme Space that deterministically romanises seven Indic scripts to ISO-15919 so Chatterbox's Latin tokeniser can process them; (2) a LoRA adapter on only the text-token predictor (Chatterbox's t3), trained on ~1,220h of licensed Indic audio with a Hindi-proxy language_id; (3) a voice-prompt recovery recipe -- an 8-11s same-language reference clip plus three sampling overrides (exaggeration 0.7, temperature 0.6, min_p 0.1; "Config B") -- that recovers commercial-class acoustic output with no acoustic-decoder training. On Hindi, the LoRA regresses accuracy and we instead use vanilla Chatterbox + Config B, giving a two-branch deployment. Evaluated on 10-utterance pilot sets with the companion PSP benchmark, Praxy Voice matches or slightly leads commercial baselines: 26.7% retroflex collapse on Telugu (vs Sarvam Bulbul 33.3%), 71% Tamil-zha collapse (vs commercial trio's 86%), 0.025 LLM-WER on Hindi (tied with Cartesia Sonic-3). For intra-sentential code-mix we add a third branch (IndicF5 + native-script transliteration) that drops code-mix LLM-WER from 0.80-0.85 to 0.14-0.27 across Hi/Te/Ta. We release R6 LoRA weights (Apache-2.0), inference code and router (MIT), and a Gradio demo.

I. Introduction · II. Related Work

Praxy Voice proposes a minimum-intervention path from frozen, non-Indic-native multilingual TTS to Indic-capable output using BUPS, text-head LoRA adaptation, and voice-prompt recovery. Its two-branch pure-script design preserves Hindi accuracy while improving Telugu and Tamil, and its related work positions romanisation-based routing and recovery sampling as distinct contributions.

I. Introduction: Production Indic TTS traditionally requires costly native-model training or per-call commercial systems, motivating a minimum-intervention wrapper around a frozen base.Native training is described as 100–1000× beyond most product teams’ reach.
I. Introduction: Praxy Voice combines BUPS, a text-head-only LoRA adapter with Hindi-proxy language_id conditioning, and BYOR plus Config B voice-prompt recovery without acoustic-decoder training.The full Chatterbox stack remains frozen apart from the text-token predictor adaptation.
I. Introduction: BUPS deterministically converts seven Brahmic scripts to ISO-15919 Latin strings, routing uncovered scripts through Chatterbox’s existing Latin-tokeniser coverage at zero additional model cost.The scripts are Devanagari, Telugu, Tamil, Kannada, Bengali, Gujarati, and Malayalam.
I. Introduction: 26.7% Telugu retroflex collapse beats Sarvam Bulbul’s 33.3%, while 71% Tamil zha collapse beats the commercial trio’s 86%; Hindi LLM-WER is 0.025, tied with Cartesia.These results place Praxy Voice in the commercial pack on all three tier-1 Indic languages, using zero commercial TTS training data.
I. Introduction: Hindi LLM-WER worsens from 0.025 to 0.334 with LoRA, so pure-script routing uses LoRA for Telugu and Tamil but vanilla Chatterbox for Hindi.Both branches share the voice-prompt and Config B inference recipe.
I. Introduction: The system adds a third code-mix branch using native-script transliteration and frozen IndicF5, alongside released MIT inference code, router, and Apache-2.0 R6 LoRA weights.The pipeline routes code-mixed inputs containing at least one Latin word of at least two characters through this branch.
II. Related Work: Chatterbox is the 23-language MIT-licensed base, contrasted with from-scratch multilingual or purpose-built Indic systems including OmniVoice, VoxCPM2, Indic Parler-TTS, and IndicF5.The related work frames Praxy Voice as complementary to from-scratch retraining rather than a replacement for it.
II. Related Work: BUPS is presented as the first known TTS routing use of lossless ISO-15919 romanisation, while the recovery contribution is the specific Config B plus same-language reference recipe rather than voice prompting itself.Prior LoRA applications target personalisation, emotion, or style; this work applies it to language extension for a base lacking the target language.

III. Method · A. BUPS: Brahmic Unified Phoneme Space

Praxy Voice uses a three-branch inference pipeline: LoRA for Telugu/Tamil, unchanged Chatterbox for Hindi, and zero-shot IndicF5 for code-mixed input. Its BUPS preprocessor makes uncovered Brahmic scripts processable by Chatterbox’s Latin tokeniser through deterministic ISO-15919 transliteration.

III. Method: The pipeline trains only the LoRA branch for pure-script Telugu and Tamil, while Hindi uses unchanged vanilla Chatterbox and code-mix uses zero-shot IndicF5.The LoRA and vanilla branches share the inference-time voice-prompt plus Config B recipe.
A. BUPS: Brahmic Unified Phoneme Space: BUPS addresses Chatterbox’s inability to tokenise raw Telugu and Tamil, which otherwise produces unsupported-language errors or byte-BPE junk tokens.Chatterbox’s MTLTokenizer covers 23 languages but does not provide useful token associations for these uncovered scripts.
A. BUPS: Brahmic Unified Phoneme Space: BUPS converts uncovered Brahmic scripts into Latin-dominant text so Chatterbox can use its dense Latin-script tokenisation paths.The method relies on Latin coverage across eight languages and preserves phonological information through romanisation.
A. BUPS: Brahmic Unified Phoneme Space: ISO-15919 provides BUPS’s deterministic, lossless mapping from Brahmic script codepoints to Latin characters with diacritics.This mapping is the phonological bridge between uncovered scripts and Chatterbox’s existing Latin tokeniser paths.
A. BUPS: Brahmic Unified Phoneme Space: The BUPS procedure therefore combines script-run segmentation, per-span transliteration, and recombination into a single preprocessing pipeline.These operations define how mixed-script input is transformed before tokenisation.
A. BUPS: Brahmic Unified Phoneme Space: BUPS first segments input into maximal single-script spans using Unicode block ranges for seven Brahmic scripts.The listed ranges cover Devanagari, Bengali, Gujarati, Tamil, Telugu, Kannada, and Malayalam.
A. BUPS: Brahmic Unified Phoneme Space: Each Brahmic span is transliterated with ISO-15919 via indic-transliteration, while Latin text, digits, and punctuation pass through unchanged.The resulting runs are concatenated into one Latin-dominant string.

B. LoRA Adaptation of the Text Head

The adaptation targets only Chatterbox’s text-token predictor with a small LoRA while freezing the acoustic generator and voice encoder. Training uses BUPS-preprocessed Indic transcripts with a Hindi-proxy language ID and approximately 1,220 hours of licensed speech, without commercial TTS training data.

Adapter scope: LoRA updates only t3 attention projections, leaving s3gen and the voice encoder frozen.The adapter uses rank 32, alpha 64, dropout 0.05, and no bias, with 7.86 M trainable parameters out of the 810 M-parameter base (0.97%).
Input format: Hindi-proxy language conditioning routes Telugu and Tamil through Chatterbox’s Hindi-supported acoustic manifold.BUPS-preprocessed transcripts are used; direct language_id=te caused ValueError and unstable training, making the Hindi-proxy path necessary.
Training data: Approximately 1,220 h of licensed Indic speech supports training, including ∼150 h Telugu, ∼700 h Hindi, and ∼280 h Tamil from Shrutilipi.The data also includes ∼20 h of Hindi emotion-labelled Rasa speech and ∼5 h per language from FLEURS; no commercial TTS training data is used.
Optimization: 8 000 steps on a single A100-80GB required ∼11 h and ∼$45, using bf16 mixed precision, AdamW, cosine scheduling, and peak LR 3 × 10−6.A quarter-LR regime with a divergence-abort heuristic stabilized training after earlier peak LR 2 × 10−5 runs diverged around step 3 600.

C. Voice-Prompt Recovery and Config B · D. Language-Specific Routing

Voice-prompt recovery supplies same-language acoustic context to a frozen decoder, while Config B sampling produces practical Telugu adaptation without acoustic-decoder training. Language-specific routing uses LoRA for Telugu and Tamil, vanilla Chatterbox for Hindi, and IndicF5 for code-mix inputs.

C. Voice-Prompt Recovery and Config B: Default Telugu inference produces intelligible but flat speech with correct words, non-native cadence, and a reported “foreigner accent.”The frozen acoustic decoder lacks a native-Indic prior despite the LoRA-adapted text head producing reasonable speech tokens.
C. Voice-Prompt Recovery and Config B: An 8–11 s same-language reference clip conditions the decoder with speaker-and-prosody information from native acoustic characteristics.The clip is supplied through Chatterbox’s audio_prompt_path interface.
C. Voice-Prompt Recovery and Config B: Config B sets exaggeration = 0.7, temperature = 0.6, and min_p = 0.1 to strengthen prosody, tighten sampling, and filter phoneme-drifting tokens.These overrides are paired with the voice prompt during inference.
C. Voice-Prompt Recovery and Config B: Voice prompting supplies native-Indic context, Config B keeps decoding on-distribution, and the LoRA text head provides inexpensive adaptation while acoustic weights remain frozen.The recipe is an inference-time substitute for acoustic-decoder adaptation.
C. Voice-Prompt Recovery and Config B: Config B was selected from a three-configuration Telugu pilot sweep with native-listener testing across declarative, interrogative, emotional, and long-narrative utterances.The authors characterize this as a practical engineering finding rather than a dense hyperparameter-optimisation result.
D. Language-Specific Routing: Pure-script deployment routes Telugu and Tamil through LoRA + BUPS with a Hindi-proxy lang_id, while Hindi uses vanilla Chatterbox without LoRA or BUPS.The pure-script branches share the voice-prompt and Config B inference recipe.
D. Language-Specific Routing: The routing rule sends lang in {te, ta} to the LoRA branch and other pure-script inputs to vanilla, with code-mix detection adding a separate branch.The code-mix branch uses native-script transliteration followed by IndicF5 for Telugu, Tamil, and Hindi.
D. Language-Specific Routing: The LoRA branch regresses Hindi semantic accuracy, whereas vanilla Chatterbox recovers commercial-class performance, making the negative control integral to routing.The deployment therefore uses two pure-script branches plus a code-mix-specific third branch.

E. Code-Mix Routing via Native-Script Transliteration · IV. Experimental Setup · A. Evaluation Benchmark

Praxy Voice routes intra-sentential code-mix through native-script transliteration and IndicF5, addressing failures caused by BUPS romanisation and Chatterbox’s single-language conditioning. Evaluation uses PSP’s phonological and corpus-level measures, interpreting per-phoneme results relatively because native Telugu audio has a substantial noise floor.

E. Code-Mix Routing via Native-Script Transliteration: Intra-sentential code-mix degrades both pure-script branches: BUPS misreads English spans, while Chatterbox’s language_id forces them through a Hindi-conditioned acoustic manifold.These failures motivate a separate routing branch for mixed-language utterances.
E. Code-Mix Routing via Native-Script Transliteration: The third branch treats code-mix as a tokeniser-input distribution problem rather than a model problem.This reframing separates code-mix handling from the LoRA contribution.
E. Code-Mix Routing via Native-Script Transliteration: IndicF5 handles code-mix zero-shot through its character-level tokeniser, without language_id, using a backbone pretrained on 1,417 h of multilingual Indic speech.No further training is used for this branch.
E. Code-Mix Routing via Native-Script Transliteration: Raw Latin spans are transliterated into native Indic script before IndicF5 inference because its Indic-only pretraining otherwise silently drops them.The transliteration restores an acoustic mapping for English spans.
E. Code-Mix Routing via Native-Script Transliteration: The branch triggers when an utterance contains a Latin alphabetic word of length ≥2; single-letter fragments and digits use the unified Indic number normaliser.The combined transliteration-to-IndicF5 path is independent of the LoRA branch.
A. Evaluation Benchmark: PSP evaluates Indic TTS with six phonological dimensions, combining per-phoneme probes with corpus-level distributional distances.The benchmark is designed so results can be interpreted without consulting the companion paper.
A. Evaluation Benchmark: Per-phoneme results are interpreted as relative system rankings because native Telugu audio exhibits ∼46% retroflex collapse, while absolute interpretation is reserved for FAD and PSD.PSP’s per-phoneme probes compare synthesised embeddings with native and non-native substitute centroids.

B. Intelligibility Metrics · C. Test Sets · D. Baselines •

The study evaluates intelligibility with multiple metrics under a shared IndicWhisper-family STT stack, using stratified 10-utterance pilot sets and code-mix smoke sets. Comparisons include three commercial APIs and three open-source systems, with Chatterbox and IndicF5 also serving deployment roles.

B. Intelligibility Metrics: Intelligibility is assessed with literal WER, LLM-WER, LLM-CER, and intent-preservation rate.LLM-WER uses a Qwen-2.5-72B semantic judge via OpenRouter.
B. Intelligibility Metrics: The same vasista22/whisper-{te,hi,ta}-large-v2 IndicWhisper-family STT stack evaluates every system for apples-to-apples comparison.This design controls for differences in STT bias across systems.
C. Test Sets: 10-utterance PSP v1 pilot sets per language are stratified across nine categories, including declarative, emotional, numerical, colloquial, and phonetically tricky speech.Each Praxy language set uses a single voice and contains n = 10 wavs.
C. Test Sets: 10-utterance code-mix smoke sets per language contain 25–35% English-token density and are stratified by tech, office, food, travel, and money topics.These sets support evaluation of the code-mix branch.
D. Baselines •: Commercial baselines are ElevenLabs v3, Cartesia Sonic-3, and Sarvam Bulbul “bulbul:v3,” accessed through public APIs using trial-tier accounts.No special credits were used.
D. Baselines •: Open-source baselines are zero-shot Indic Parler-TTS, vanilla Chatterbox without LoRA or BUPS, and zero-shot AI4Bharat IndicF5.These systems are evaluated on the same smoke sets.
D. Baselines •: Vanilla Chatterbox is also the Hindi-branch deployment model, while IndicF5 is the back-end of the code-mix branch.Chatterbox serves as both the study’s internal reference and a deployment model.

V. Results · A. Headline Results · B. Config B Ablation

Praxy matches or exceeds commercial baselines on key Telugu, Tamil, and Hindi phonological or intelligibility measures, while Hindi FAD remains its main weakness. Config B is selected because it dominates the Telugu ablation and listener comparisons across measured axes and qualitative categories.

A. Headline Results: 26.7% retroflex collapse on Telugu beats Sarvam Bulbul’s 33.3%, while 71% Tamil-zha collapse improves on the commercial trio’s 86%.Both comparisons use small pilot token counts: n = 15 retroflex tokens for Telugu and n = 7 zha tokens for Tamil.
A. Headline Results: The headline evaluation uses 10 utterances per Praxy row and 20 per commercial row, with Telugu and Tamil using the LoRA branch and Hindi using the vanilla branch.Praxy systems use the voice-prompt plus Config B recipe; Tamil-zha scoring applies only to Tamil.
A. Headline Results: On Tamil, Praxy matches Sarvam on PSD at 71.2 versus 72.3, while Indic Parler remains best on RR and LF.The Tamil-zha comparison is 71% for Praxy versus 86% for the commercial trio.
A. Headline Results: On Hindi, Praxy-vanilla ties Cartesia on LLM-WER and reaches perfect intent-preservation, but its FAD of 439 trails Sarvam at 212 and Cartesia at 267.The paper identifies Hindi FAD as the single axis where acoustic-decoder adaptation would measurably help, although that adaptation is out of scope.
B. Config B Ablation: The Telugu ablation sweeps three sampling configurations while holding the R6 LoRA branch and Cartesia-Te-male reference audio fixed.The pilot set contains n = 10 utterances.
B. Config B Ablation: Config B uses exaggeration 0.7, temperature 0.6, and min_p 0.1, whereas Config A uses repetition_penalty 1.2 and min_p 0.03.Config A is intended to preserve word-final syllables through a lower repetition penalty.
B. Config B Ablation: Config B dominates Config A and Config C on LLM-WER, intent-preservation, and FAD, achieving 5× better LLM-WER than A and 2× better than C.Config A breaks speech coherence rather than fixing word-final syllables, while Config C produces partial-fidelity output through under-conditioning.
B. Config B Ablation: Listeners placed Config B unambiguously first across declarative, interrogative, emotional, and long-narrative categories, informing its selection without constituting a formal MOS result.This parallel native-Telugu listener ear test is reported as qualitative evidence only.

C. Scope-of-Method Control: Hindi With vs Without LoRA · D. Reference-Audio Source Ablation

The Hindi control shows that LoRA+BUPS is harmful on Hindi, so vanilla Chatterbox remains the intended Hindi path. Telugu reference ablation shows that same-language voice prompts improve native-like output, whereas an English prompt degrades FAD.

C. Scope-of-Method Control: Hindi With vs Without LoRA: Table III compares R6 LoRA+BUPS for Telugu/Tamil with vanilla Chatterbox for Hindi under the same Config B and Cartesia-Hindi reference.The comparison isolates the model-variant choice while holding the reference and sampling configuration fixed.
C. Scope-of-Method Control: Hindi With vs Without LoRA: 13× worse LLM-WER occurs with LoRA+BUPS than with vanilla Chatterbox on Hindi.This supports disabling the LoRA branch for the intended Hindi path.
C. Scope-of-Method Control: Hindi With vs Without LoRA: 40% of the Hindi LLM-WER gap is recovered by disabling BUPS, leaving the LoRA branch substantially worse than vanilla Chatterbox.The passage attributes the remaining degradation to the LoRA adapter’s interaction with Hindi processing.
C. Scope-of-Method Control: Hindi With vs Without LoRA: The Hindi degradation is consistent with training Telugu/Tamil BUPS-romanised text under Hindi language_id, then applying the LoRA to Devanagari Hindi.The text head therefore expects romanised input, corrupting the native Hindi tokeniser path.
D. Reference-Audio Source Ablation: Four reference-audio sources were swept on Telugu with the LoRA branch and Config B fixed, varying only the voice-prompt reference.This isolates reference-audio source as the ablated factor.
D. Reference-Audio Source Ablation: 26% FAD degradation occurs with the 49-second English memo, while same-language Sarvam-Te and Cartesia-Te references win on the Telugu pilot.The ablation used n = 10 and held the Praxy R6 LoRA branch plus Config B sampling fixed.
D. Reference-Audio Source Ablation: Same-language Telugu references close the FAD-and-PSD gap to native dramatically, whereas cross-language English prompts harm native-like output.The proposed explanation is that voice prompts encode both speaker timbre and prosody, with cross-language prompts shifting prosody toward English.

E. R5 →R6 Training-Scale Delta … A. Why voice-prompt recovery works (and why only sometimes)

Scaling from R5 to R6 substantially improved fidelity and intelligibility but worsened prosodic similarity, motivating voice-prompt recovery. The code-mix branch benefited strongly from native-script transliteration, while the frozen acoustic decoder remained effective when conditioned by same-language voice prompts.

E. R5 →R6 Training-Scale Delta: LLM-WER improved 5× from 0.171 to 0.034 after scaling from R5’s 85 h to R6’s ∼1,220 h multilingual mix.The scaleup included Shrutilipi and a Te-dominant R5 mix.
E. R5 →R6 Training-Scale Delta: FAD fell 34% from 534 to 355, while PSD rose 338% from 14 to 62.The passage attributes the PSD regression to prosodic drift as the token path broadened.
E. R5 →R6 Training-Scale Delta: Retroflex collapse remained 40% →40% at R6-no-ref, consistent with LoRA-on-t3 not changing acoustic-decoder discrimination.The result supports separating text-token prediction improvements from acoustic-decoder phonological discrimination.
F. Code-Mix Branch: The code-mix experiment used 10 utterances per language with 25–35% English-token density, comparing IndicF5 with and without native-script transliteration.Commercial reference points used the same commercial trio as the headline results.
F. Code-Mix Branch: 76% relative LLM-WER reduction on Hindi (0.855 →0.198) and 82% on Telugu (0.798 →0.142) accompanied intent-preservation increases from ≤10% to 70–80%.Tamil improved less, with a 64% relative WER reduction and 60% intent preservation.
F. Code-Mix Branch: Tamil’s smaller improvement was consistent with IndicF5’s 80 h Tamil pre-training subset being the smallest evaluated-language subset.The passage reports Tamil at 64% relative WER reduction and 60% intent preservation.
A. Why voice-prompt recovery works (and why only sometimes): A native same-language voice prompt pulls frozen s3gen acoustic output onto the native manifold after LoRA-adapted t3 emits on-manifold speech token sequences.The acoustic decoder converts tokens to mel and audio while conditioning on text-adapted token embeddings and the voice-prompt embedding.

B. Code-mix evaluation gap · C. PSP as a diagnostic loop · D. Limitations •

The paper identifies a code-mix evaluation artefact, shows how PSP localises errors to guide inference-time intervention, and reports limitations spanning pilot scale, acoustic adaptation, MOS, Hindi FAD, and reference-audio use.

B. Code-mix evaluation gap: Cartesia’s near-zero Hi code-mix LLM-WER reflects American-English pronunciation recognised by Whisperlarge-v3 STT, whereas the recipe intentionally produces Indianised English pronunciation.The native-script transliteration is designed to match conversational Indian code-switching but is penalised by the evaluation setup.
C. PSP as a diagnostic loop: PSP’s per-dimension decomposition provides a diagnostic loop that connects observed failures to targeted intervention choices.The paper presents the sequence of §V results as an example of PSP’s advocated usage pattern.
C. PSP as a diagnostic loop: 534 →355 Telugu FAD after R5→R6 scaleup exposed a 14 →62 PSD gap, localising the remaining problem to prosodic conditioning and motivating voice-prompt recovery + Config B.The inference-time fix closed PSD 4.7× to 13.1.
C. PSP as a diagnostic loop: 13× Hindi LLM-WER regression with per-phoneme cells remaining at 0% localised the LoRA failure to an intervening factor rather than phonological accuracy.This contrasts with the Telugu localisation, where the remaining issue was prosodic conditioning.
D. Limitations •: n = 10 pilot utterances and 15–39 retroflex tokens per Telugu/Tamil cell cannot statistically separate 5-percentage-point differences.300-utterance full benchmarks are in progress for v2.
D. Limitations •: Acoustic-decoder adaptation remains unexplored because s3gen training did not fit meaningfully on A100-80GB, while batch-size-1 training was estimated at 64+ days for 4000 steps.H100-80GB or larger would unblock this compute-budget limitation.
D. Limitations •: No formal MOS panels have been run; native-listener ear tests guided ablations, while Karya MOS calibration at 300-utterance scale is deferred to v2.The ear test especially informed Config B.
D. Limitations •: 439 vs Sarvam 212 and Cartesia 267 Hindi FAD remains moderate despite Praxy Hi vanilla tying Cartesia on WER, marking acoustic adaptation as the measurable opportunity.Hindi FAD is identified as the one axis where acoustic adaptation would help.

VII. Release • · VIII. Conclusion

The paper releases the Praxy Voice weights, implementation stack, demo, and evaluation artefacts, and concludes that a frozen non-Indic-native base can reach commercial-class output through BUPS, text-head LoRA, and Config B voice-prompt recovery. On Hindi, the LoRA regresses semantic accuracy, motivating the voice-prompt-plus-Config-B approach described in the conclusion.

VII. Release •: R6 LoRA weights from step 8000 are released under Apache-2.0.The weights are hosted at the Praxy Voice Hugging Face repository.
VII. Release •: Inference code, BUPS, Config B, the language router, and the unified Indic number/date/currency normaliser are released under MIT.These components are available in the Praxy GitHub repository.
VII. Release •: A Gradio demo supports BYOR voice cloning for Telugu, Tamil, and Hindi.The demo is available through Hugging Face Spaces, with the link provided in the Hugging Face repository README.
VII. Release •: Evaluation scorecards and benchmark artefacts are bundled in the PSP companion paper’s reproducibility JSON.This release provides the accompanying evaluation materials rather than only model and inference files.
VIII. Conclusion: A frozen, non-Indic-native multilingual TTS base reaches commercial-class output on three tier-1 Indic languages through a minimum-intervention recipe.The recipe combines BUPS ISO-15919 romanisation, a LoRA adapter on only the text head, and an inference-time voice-prompt recovery recipe with Config B sampling overrides.
VIII. Conclusion: BUPS ISO-15919 romanisation addresses uncovered Brahmic scripts, while the LoRA adapter is applied only to the text head.The conclusion presents these as two components of the minimum-intervention recipe, alongside Config B voice-prompt recovery.
VIII. Conclusion: On Hindi, the natively covered language, the LoRA adapter actively regresses semantic accuracy.The supplied conclusion passage states that the same voice-prompt-plus-Config-B recipe is applied, but the passage ends before specifying the full outcome.