Praxy Voice: Modular Indic TTS Adaptation

• Praxy Voice is an open-source framework that leverages deterministic romanisation (BUPS) and minimal LoRA adaptation to enable commercial-class Indic TTS from a frozen non-Indic base.
• It employs a three-branch architecture with tailored configurations, achieving competitive phonological and acoustic performance for Telugu, Tamil, and Hindi synthesis.
• The design emphasizes practical deployment by integrating text-side adaptation, inference-time voice-prompt recovery, and code-mix routing to address multilingual TTS challenges.

Searching arXiv for the named topic and closely related work to ground the article in current papers. Searching for “Praxy Voice” and the cited arXiv ids to verify metadata and related papers. Praxy Voice most specifically denotes an open-source recipe for producing commercial-class Telugu, Tamil, and Hindi text-to-speech from a frozen non-Indic base model, ResembleAI Chatterbox, without training a new acoustic decoder and without any commercial TTS training data. The recipe combines BUPS, a Brahmic Unified Phoneme Space that romanises seven Indic scripts to ISO-15919, a LoRA adapter on Chatterbox’s text-token predictor $t_3$, and a voice-prompt recovery configuration built around an 8–11 s same-language reference clip and the sampling overrides exaggeration $0.7$, temperature $0.6$, and min_p $0.1$ (“Config B”) (Menta, 28 Apr 2026). In adjacent work, the same label is also used interpretively for practical multi-identity synthetic voice, low-resource few-shot cloning, low-latency voice-to-voice agents, and controllable perceptual voice-quality manipulation, so the term now sits at the intersection of multilingual TTS adaptation, speaker conditioning, security, and voice UX (Pani et al., 2023, Karki et al., 2024, Shrestha et al., 26 Jan 2026, Purwar et al., 25 Sep 2025, Rautenberg et al., 7 Nov 2025).

1. Scope and conceptual range

Praxy Voice, in its narrowest sense, is the 2026 system titled “Praxy Voice: Voice-Prompt Recovery + BUPS for Commercial-Class Indic TTS from a Frozen Non-Indic Base at Zero Commercial-Training-Data Cost,” whose target is commercial-class Telugu, Tamil, and Hindi synthesis from a frozen multilingual base that does not natively tokenise Telugu or Tamil (Menta, 28 Apr 2026). Its explicit motivation is the gap between commercial systems and the strongest open-source bases: Chatterbox does not tokenise Telugu or Tamil, while Indic Parler-TTS and IndicF5 are from-scratch Indic models rather than lightweight adaptations of a frozen non-Indic base.

In the surrounding literature, “Praxy Voice” is also used as an interpretive label for several other practical voice configurations. One strand uses the term for a “practical multi-identity synthetic voice” in the setting of Voice Identity Morphing, where one synthetic utterance can match two enrolled speakers in automatic speaker verification (Pani et al., 2023). Another uses it as shorthand for low-resource, few-shot Nepali voice cloning built from a speaker encoder, a Tacotron2-style synthesizer, and a neural vocoder (Karki et al., 2024, Shrestha et al., 26 Jan 2026). Additional adjacent work connects the idea to low-latency voice-to-voice agents, controllable expressiveness, perceptual voice-quality control, and qualitative voice UX research (Purwar et al., 25 Sep 2025, Neekhara et al., 2021, Narain et al., 27 May 2025, Seaborn et al., 2024). This suggests that the term has become less a single architecture than a practical design space for voice systems that prioritize deployment constraints, controllability, and task-specific robustness.

2. Core architecture of the 2026 Praxy Voice system

The 2026 Praxy Voice system is a three-branch deployment wrapped around Chatterbox. Its central claim is that the minimum effective intervention is not acoustic-decoder retraining, but text-side adaptation plus inference-time acoustic steering. Chatterbox’s $t_3$ text head is adapted with LoRA, while the acoustic decoder $s_3\text{gen}$ and voice encoder $\text{ve}$ remain frozen (Menta, 28 Apr 2026).

BUPS is the enabling preprocessing layer. It segments input into maximal script runs, applies ISO-15919 transliteration to Brahmic spans from Devanagari, Bengali, Gujarati, Tamil, Telugu, Kannada, and Malayalam, and leaves non-Brahmic spans untouched. The purpose is not phonological abstraction but deterministic romanisation: retroflexes, aspirates, nukta forms, and vowel distinctions are preserved in Latin with diacritics so that Chatterbox’s Latin-capable tokenizer can process Indic input. Telugu- and Tamil-script inputs are therefore converted into a token stream that the frozen multilingual base can consume without changing its tokenizer (Menta, 28 Apr 2026).

The LoRA adapter is attached only to the attention projections $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$ inside $t_3$. With rank $r=32$, LoRA $0.7$0, dropout $0.7$1, and no bias, it introduces $0.7$2M trainable parameters, about $0.7$3 of the $0.7$4M-parameter base. Training uses approximately $0.7$5 hours of licensed Indic audio from IndicTTS, Rasa, FLEURS, and Shrutilipi, with Telugu and Tamil transcripts BUPS-romanised and forced through the Hindi proxy language_id=hi. Optimization uses AdamW, bf16 mixed precision, batch size $0.7$6, gradient clipping $0.7$7, a cosine schedule with $0.7$8-step warmup, and a peak learning rate of $0.7$9, finishing at $0.6$0 steps on a single A100-80GB in roughly $0.6$1 hours (Menta, 28 Apr 2026).

The resulting deployment logic is summarized below.

Branch	Trigger	Core path
Telugu/Tamil pure-script	Pure te or ta native script	BUPS $0.6$2 LoRA-adapted $0.6$3 $0.6$4 frozen Chatterbox acoustic decoder with same-language reference clip + Config B
Hindi pure-script	Pure Devanagari	Vanilla Chatterbox $0.6$5 same-language reference clip + Config B
Code-mix	Any Latin word of length $0.6$6	Native-script transliteration via Claude Haiku 4.5 $0.6$7 IndicF5

The voice-prompt recovery step is the acoustic half of the recipe. After BUPS and LoRA, default Chatterbox decoding was reported as intelligible but flat and non-native. Praxy Voice therefore adds an 8–11 s same-language reference clip through audio_prompt_path, allowing the frozen voice encoder to steer the decoder toward a same-language acoustic manifold, and couples it with Config B: exaggeration $0.6$8, temperature $0.6$9, and min_p $0.1$0 (Menta, 28 Apr 2026).

3. Evaluation, routing behavior, and empirical performance

Praxy Voice is evaluated with the companion PSP benchmark, whose phonological dimensions include retroflex collapse rate (RR), Tamil “ழ” collapse rate (ZF), aspiration fidelity (AF), length fidelity (LF), Fréchet Audio Distance (FAD), and Prosodic Signature Divergence (PSD). Intelligibility is measured with literal WER, LLM-WER, LLM-CER, and intent preservation. The reported test bed is a 10-utterance pilot set per language, with separate 10-utterance code-mix smoke sets for Hindi, Telugu, and Tamil (Menta, 28 Apr 2026).

On Telugu, Praxy Voice reports $0.1$1 retroflex collapse, compared with Sarvam Bulbul’s $0.1$2, Cartesia Sonic-3’s $0.1$3, ElevenLabs’ $0.1$4, and Indic Parler-TTS’s $0.1$5. Its FAD is $0.1$6, against Sarvam’s $0.1$7, Indic Parler-TTS’s $0.1$8, ElevenLabs’ $0.1$9, and Cartesia’s $t_3$0. LLM-WER is $t_3$1 with intent $t_3$2, essentially in the same band as Sarvam and Cartesia at $t_3$3 (Menta, 28 Apr 2026).

On Tamil, Praxy Voice reports $t_3$4 Tamil “ழ” collapse, versus $t_3$5 for Sarvam Bulbul, Cartesia Sonic-3, and ElevenLabs; Indic Parler-TTS is lower at $t_3$6. Praxy Voice’s Tamil RR is $t_3$7, with FAD $t_3$8, PSD $t_3$9, LLM-WER $s_3\text{gen}$0, and intent $s_3\text{gen}$1. On Hindi, the paper reports that the LoRA branch regresses accuracy and is therefore disabled: R6 LoRA + BUPS yields LLM-WER $s_3\text{gen}$2, R6 LoRA without BUPS yields $s_3\text{gen}$3, while vanilla Chatterbox plus Config B yields $s_3\text{gen}$4 LLM-WER and $s_3\text{gen}$5 intent, tying Cartesia Sonic-3’s $s_3\text{gen}$6 but trailing Sarvam and ElevenLabs on acoustic distance (Menta, 28 Apr 2026).

The code-mix branch uses native-script transliteration plus frozen IndicF5. Raw IndicF5 on code-mixed input yields LLM-WER $s_3\text{gen}$7 for Hindi, $s_3\text{gen}$8 for Telugu, and $s_3\text{gen}$9 for Tamil, with intent $\text{ve}$0, $\text{ve}$1, and $\text{ve}$2, respectively. The Praxy code-mix branch reduces these to $\text{ve}$3, $\text{ve}$4, and $\text{ve}$5, with intent $\text{ve}$6, $\text{ve}$7, and $\text{ve}$8. This suggests that, within the reported setting, routing and transliteration are at least as consequential as model choice for intra-sentential Indic–English synthesis (Menta, 28 Apr 2026).

Ablations also identify Config B as the decisive inference setting. For Telugu with R6 LoRA and a Cartesia-Telugu reference, Config A gives LLM-WER $\text{ve}$9, intent $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$0, FAD $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$1, PSD $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$2; Config C gives $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$3, $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$4, $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$5, $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$6; Config B gives $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$7, $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$8, $$q_{\text{proj}}, k_{\text{proj}}, v_{\text{proj}}, o_{\text{proj}}$$9, $t_3$0. No formal MOS is reported; selection was guided by a native-Telugu ear test and task metrics (Menta, 28 Apr 2026).

4. Relation to few-shot cloning and expressive synthesis

Praxy Voice belongs to a broader technical lineage in which a speaker-conditioned text-to-speech stack separates identity encoding, acoustic generation, and waveform synthesis. Low-resource Nepali systems described in the same research corpus use the canonical three-model structure—speaker encoder, Tacotron2-style synthesizer, and WaveNet or WaveRNN vocoder—with speaker embeddings computed from a few audio samples and reused at inference for unseen speakers. One such system reports MOS $t_3$1 for naturalness and $t_3$2 for similarity, with PESQ $t_3$3 on validation and $t_3$4 on test; another reports overall MOS quality $t_3$5, MOS similarity $t_3$6, and mean cosine similarity between original and cloned speaker embeddings of about $t_3$7 (Karki et al., 2024, Shrestha et al., 26 Jan 2026).

Accent-aware cloning work for Indian and Western accents follows a similar encoder–synthesizer–vocoder pattern, but with explicit multi-accent data balancing and post-vocoder noise reduction. The reported speaker encoder is trained on $t_3$8 hours from $t_3$9 speakers, with $r=32$0 hours of Indian accents and $r=32$1 hours of Western accents, and the system is evaluated with MOS, Gross Pitch Error, and Spectral Distortion on unseen Indian and Western speakers (R et al., 2024). This suggests that Praxy Voice’s reliance on a same-language reference clip is one pragmatic alternative to full multi-accent acoustic retraining.

Expressive cloning extends the same decomposition by adding style variables. “Expressive Neural Voice Cloning” conditions a Tacotron-2-like synthesizer not only on a speaker embedding $r=32$2, but also on explicit pitch contour $r=32$3, latent style tokens $r=32$4, and rhythm $r=32$5, enabling style transfer, text-driven synthesis, and fine-grained style control for unseen speakers. Relative to a Tacotron2+GST baseline, the proposed model reduces Gross Pitch Error from $r=32$6 to $r=32$7 in zero-shot imitation and improves naturalness MOS from $r=32$8 to $r=32$9 on text synthesis (Neekhara et al., 2021). Compared with such factorized style-control systems, Praxy Voice is narrower: it does not expose explicit prosody or style variables, but instead uses voice-prompt recovery to bias a frozen acoustic decoder toward a language-appropriate manifold.

5. Security, identity, and attack surfaces

A recurring theme in adjacent work is that practical voice systems create not only assistive and generative capabilities, but also new biometric and communication vulnerabilities. Voice Identity Morphing defines a voice-based morph attack in which two speaker embeddings are fused—specifically by simple averaging, $0.7$00—and then conditioned through DeepTalk, Tacotron 2, and WaveRNN to synthesize a single utterance that matches both contributing speakers in automatic speaker verification (Pani et al., 2023).

On Librispeech, this attack achieves morph-sample MMPMR $0.7$01 on ECAPA-TDNN and $0.7$02 on x-vector at FMR $0.7$03; at FMR $0.7$04, speaker-pair MMPMR remains $0.7$05 for ECAPA-TDNN and $0.7$06 for x-vector. The paper explicitly frames this as a breach of the standard biometric assumption that one biometric sample corresponds to one identity (Pani et al., 2023). A plausible implication is that any practical voice platform that supports high-fidelity speaker conditioning, including systems adjacent to Praxy Voice, inherits a dual-use profile: the same abstractions that enable flexible identity transfer for benign synthesis also lower the barrier to multi-identity or impersonation attacks.

A distinct security line concerns secure transmission rather than biometric spoofing. Data-over-Voice systems compress speech with Codec2, encrypt the compressed bitstream with AES-256 in CTR mode, map it to pseudo-speech audio using a codebook of short harmonic waveforms, and then transmit it over 3G or VoIP channels. The reported system reaches about $0.7$07 kbps raw DoV over 3G and up to $0.7$08 kbps over VoIP, with effective secure voice around $0.7$09 or $0.7$10 kbps depending on mode (Krasnowski et al., 2021). This is orthogonal to Praxy Voice’s synthesis agenda, but it underscores that “practical voice” research spans both generative quality and transport-layer privacy.

6. Real-time operation, interpretability, and voice UX

Praxy Voice is not a real-time voice-to-voice paper, but adjacent work on low-latency agents shows the systems problem that emerges once high-quality TTS is inserted into a conversational loop. The i-LAVA architecture uses Silero-VAD, Whisper v3-large-turbo, gpt-4o-mini, and CSM-1B with Mimi/RVQ audio tokenization. In streaming mode with 16 RVQ iterations, GPU first-chunk latency is $0.7$11 ms and TTS RTF is $0.7$12; on a separate GPU sweep, 16 iterations yield first chunk $0.7$13 ms and RTF $0.7$14, while 32 iterations yield $0.7$15 ms and $0.7$16 (Purwar et al., 25 Sep 2025). This suggests that if Praxy Voice were embedded in a live agent, decoder-side sampling and vocoder choices would become as important as linguistic coverage.

Interpretability is addressed in a different way by work on perceptual voice qualities. One paper models seven dimensions—intelligibility, imprecise consonants, harsh voice, naturalness, monoloudness, monopitch, and breathiness—using frozen SSL embeddings and linear probes, with average Spearman $0.7$17 up to $0.7$18 and average AUC around $0.7$19–$0.7$20 on SAP, plus strong zero-shot transfer to English and Italian atypical speech and qualitative transfer to affective speech (Narain et al., 27 May 2025). Another shows that a YourTTS/VITS-based system augmented with a Conditional Continuous Normalizing Flow can manipulate a global creak probability and still induce localized changes in voiced segments, with perceived creak rising from $0.7$21 in the suppressed condition to $0.7$22 in the amplified condition, while MOS remains between $0.7$23 and $0.7$24 for synthetic speech (Rautenberg et al., 7 Nov 2025). These studies indicate a direction that Praxy Voice does not yet implement: explicit, interpretable control over phonation and speaking style.

Finally, qualitative Voice UX research expands the frame beyond tokenization, WER, and FAD. A systematic review of $0.7$25 studies reports that interviews appear in $0.7$26 of papers, observation and field studies in about $0.7$27, and that voice UX must be analyzed as a social, embodied, and time-bound phenomenon rather than only as a recognition or synthesis problem (Seaborn et al., 2024). For Praxy Voice, this implies that commercial-class output is only one layer of system quality. Reference-audio collection, multilingual routing, bystander privacy, code-mix expectations, and speaker identity handling are equally part of the operational definition of a practical voice system.

7. Limitations and significance

Praxy Voice’s central limitation is that it is not a universal single-model solution. Hindi improves only when LoRA is disabled; Telugu and Tamil improve only when BUPS, the Hindi proxy route, a same-language reference clip, and Config B are combined; code-mix requires an entirely separate IndicF5 branch with native-script transliteration (Menta, 28 Apr 2026). The paper is explicit that acoustic-decoder adaptation was not attempted to completion because $0.7$28 fine-tuning did not fit within A100-80GB constraints at a practical batch size, and that no formal MOS is yet reported.

The evaluation itself is also intentionally preliminary. The PSP pilot uses $0.7$29 utterances per language, and the paper notes that per-phoneme probes have a non-trivial native noise floor—for example, calibration yields about $0.7$30 retroflex collapse even in native Telugu recordings—so RR and ZF should be treated as relative rather than absolute. Future work identified in the paper includes larger PSP v2 benchmarks, formal MOS, broader BUPS coverage, acoustic-decoder LoRA with stronger hardware, and code-mix fine-tuning rather than zero-shot transliteration (Menta, 28 Apr 2026).

Even with those caveats, the significance of Praxy Voice is clear. It demonstrates that commercial-class Indic TTS need not require a new tokenizer, a new acoustic decoder, proprietary training audio, or from-scratch multilingual retraining. Instead, deterministic script routing, a $0.7$31M-parameter LoRA on the text head, and inference-time voice-prompt recovery are sufficient to place Telugu and Tamil in the commercial range on the reported pilot metrics, while Hindi is best served by a separate vanilla branch and code-mix by a third router branch (Menta, 28 Apr 2026). In that sense, Praxy Voice is best understood not as a monolithic model, but as a modular recipe for turning existing voice components into a deployable multilingual system under strict data, compute, and licensing constraints.