Hinglish TTS

Open-source text-to-speech for Hindi-English code-mixed ("Hinglish") speech. Built on IndicF5 (AI4Bharat, 330M params) with a two-line duration patch and automatic script normalisation via IndicXlit. Accepts Devanagari, Roman-script Hindi, mixed-script sentences, and English with Indian named entities — all in the same API call. 4.70 / 5.0 mean intelligibility (30-sentence eval set, three-ASR consensus).

---

Sample outputs

The 6 inputs below come from the rubric v2.1 production run (full table and per-sentence scores in EVALUATION_REPORT.md). Live demo with all 30 sentences: hinglish-tts.vercel.app.

Input	Category	Score
कल मुझे दिल्ली जाना है।	Pure Devanagari	5/5
yaar tu kal kya kar raha tha	Pure Roman Hinglish	5/5
kal mujhe office jaana hai	Pure Roman (English loan)	5/5
Boss को बता देना kal मैं leave पर रहूँगा	Mixed script	5/5
Mera presentation tomorrow है, और मैं nervous हूं।	Mixed script	5/5
My friend Aishwarya from Chennai is visiting Bengaluru	English with NE	4/5

---

Quality numbers

Overall

Metric	Value
Mean intelligibility	4.70 / 5.0
Silence or skip	0 / 30
Sentences scoring 5/5	19 / 30
Sentences scoring ≤ 2	0 / 30

Scoring rubric: character error rate (CER) after Devanagari normalisation, three-ASR consensus (AssemblyAI + Deepgram + Groq Whisper), 30-sentence eval set. Naturalness is evaluated by ear only — see Naturalness.

By input category

Category	n	Score	What it covers
pure_devanagari	8	4.62	Standard Hindi in Devanagari script
pure_roman	8	4.75	Colloquial Hinglish in Roman script
mixed_script	9	4.88	Mid-sentence Devanagari ↔ Roman switching
english_with_NE	6	4.50	English sentences with Indian proper nouns

Comparison to other open models

Same eval set, same rubric.

Model	Overall	pure_roman	mixed	eng_NE	pure_dev
This package	4.70	4.75	4.88	4.50	4.62
Kokoro v1.0 (Hindi)	3.90	2.50	3.88	4.83	4.62
Indic Parler-TTS	3.40	1.75	3.00	4.67	4.50
IndicF5 (unpatched, no preprocessing)	2.13	1.00	1.62	1.00	4.62
SPRINGLab F5-Hindi	2.07	1.00	1.38	1.00	4.62

The gap is largest on pure_roman (+2.25 over Kokoro), the dominant input register for Indian chat and voice-interface applications. Kokoro leads on english_with_NE — see Known limitations.

Full methodology and per-sentence breakdown: EVALUATION_REPORT.md.

---

Installation

Prerequisites: Python 3.10–3.11, CUDA GPU (≥ 6 GB VRAM), a HuggingFace account with model gating accepted.

# 1. Clone
git clone https://github.com/harrrshall/hinglish-tts.git && cd hinglish-tts

# 2. Install IndicF5 and inference dependencies
pip install git+https://github.com/AI4Bharat/IndicF5.git \
    "transformers==4.49.0" "accelerate==0.33.0" \
    "numpy>=2.0,<2.1" soundfile

# 3. Accept model gating — visit the link below and click "Agree and access repository"
#    https://huggingface.co/ai4bharat/IndicF5

# 4. Set your HuggingFace token
export HF_TOKEN=hf_your_token_here

# 5. Install preprocessing (requires Python ≤ 3.11; skip for Devanagari-only input)
pip install ai4bharat-transliteration

Duration patch: inference.py automatically patches f5_tts/infer/utils_infer.py in your site-packages on first import. This is required — the unpatched model truncates all Roman-script output. Run python inference.py dummy --patch-status to verify it is active.

---

Usage

Python API

from inference import load_model, synthesize
import soundfile as sf

model = load_model()   # reads HF_TOKEN from environment; downloads weights once (~1.3 GB)

REF_AUDIO = "data/reference_audio/hindi_ref.wav"   # any 3–10s mono 24 kHz Hindi clip
REF_TEXT  = open("data/reference_audio/hindi_ref.txt").read().strip()

Pure Devanagari

audio = synthesize(model, "कल मुझे दिल्ली जाना है।", REF_AUDIO, REF_TEXT)
sf.write("out.wav", audio, 24000)

Roman-script Hinglish — preprocessing converts to Devanagari automatically

audio = synthesize(model, "yaar tu kal kya kar raha tha", REF_AUDIO, REF_TEXT)

Mixed script — mid-sentence script switching

audio = synthesize(model,
    "Boss को बता देना kal मैं leave पर रहूँगा, kuch personal काम है।",
    REF_AUDIO, REF_TEXT)

English with Indian named entities

audio = synthesize(model,
    "My friend Aishwarya from Chennai is visiting Bengaluru next week.",
    REF_AUDIO, REF_TEXT)

Raw Devanagari, no preprocessing — for phonetic experiments or pre-normalised input

audio = synthesize(model, "मेरा नाम ज़ारा है।", REF_AUDIO, REF_TEXT, preprocess=False)

Command line

python inference.py "yaar tu kal kya kar raha tha" \
    --ref-audio data/reference_audio/hindi_ref.wav \
    --ref-text "$(cat data/reference_audio/hindi_ref.txt)" \
    --out out.wav

# Check duration patch status before a batch run
python inference.py placeholder --patch-status \
    --ref-audio x --ref-text x --out /dev/null

---

How it works

1. Preprocessing (scoring/scripts/lib_normalize.py) Roman tokens are transliterated to Devanagari by IndicXlit. Two override tables correct cases IndicXlit gets wrong: 36 English loanwords with fixed canonical Devanagari forms (office → ऑफिस, laptop → लैपटॉप, party → पार्टी), and 4 short Hindi function words that IndicXlit misreads as English phonetics (mai → मैं, tu → तू, aa → आ, hu → हूं). Pure Devanagari input passes through unchanged. The function is idempotent.

2. Duration patch IndicF5 v12 allocates synthesis canvas time using UTF-8 byte counts. Devanagari encodes as ~3 bytes/character; ASCII as 1. For Roman-script input this allocates ~3× less canvas than needed and truncates the audio — the "Mode A" failure that produces 21/30 silent outputs on the unpatched model. inference.py patches utils_infer.py in site-packages at import time, replacing byte counts with non-whitespace character counts. The patch is idempotent and emits a DEBUG-PATCH: diagnostic line per call confirming it is active.

3. IndicF5 inference IndicF5 is a 330M-parameter flow-matching TTS model from AI4Bharat, trained on Indic-script audio. It performs zero-shot voice cloning: provide a 3–10s reference clip and its transcript, and the model generates speech in that voice. No fine-tuning was applied to reach the 4.70 score — the gain is entirely from preprocessing and the duration patch. Output: 24 kHz mono PCM float32.

---

Known limitations

Summarised here; detailed with examples and architectural notes at KNOWN_LIMITATIONS.md.

• Synthetic texture. Outputs are intelligible and fluent but audibly synthetic. A trained ear identifies every clip as TTS regardless of score.

• Prosodic flatness. Long sentences, questions, and exclamations receive similar flat delivery. Ellipsis and ! produce subtle pacing and pitch effects but not dramatically different emphasis.

• No naturalness number. The 4.70 is intelligibility only. Automatic naturalness predictors (UTMOS, SQUIM_MOS) invert directionally on Hindi by ~2 ranks and cannot be trusted here.

• Whitelist generalisation (n=30). The English-loan and Indian-NE whitelists were tuned to the 30-sentence eval vocabulary. Unknown tokens fall through to IndicXlit, which handles common cases well but degrades on ambiguous short tokens and uncommon proper nouns.

• Kokoro leads on english_with_NE. Kokoro 4.83 vs 4.50. This package transliterates English sentences to Hindi-accented phonetics ("माय फ्रेंड ऐश्वर्या"); listeners who prefer English-mode proper-noun rendering should evaluate Kokoro for this category.

• Gated model weights. IndicF5 requires HuggingFace account and manual gating acceptance. Weights are not redistributed here.

• Python ≤ 3.11 for Roman-script preprocessing. IndicXlit depends on fairseq, which has a dataclass incompatibility with Python 3.12. Whitelist paths (covering the eval-set vocabulary) work on any Python version; only IndicXlit fallback requires 3.11 or earlier.

---

Naturalness

Not auto-scored. A native Hindi speaker evaluated all 30 v2.1 outputs. Summary:

• Pure Hindi and normalised Hinglish are fluent and natural for conversational sentences. Vowel lengths, aspiration, and basic prosodic contours are correct.
• Delivery is flat on longer sentences. Output is identifiably synthetic even at 5/5 intelligibility.
• English loanwords and Indian names are produced with consistent Hindi phonetics — "Bengaluru" with Indian vowels, not anglicised.
• The model responds to fine Devanagari distinctions: vowel length (ि vs ी), aspiration (ख vs क), and nukta (ज़ vs ज) produce measurably different acoustic output when input is correctly marked up.

---

Citation

If you use the eval set, scoring rubric, or preprocessing code from this work:

@misc{singh2026hinglish,
  author       = {Harshal Singh},
  title        = {Hinglish TTS: IndicF5 with IndicXlit Preprocessing},
  year         = {2026},
  howpublished = {\url{https://github.com/harrrshall/hinglish-tts}},
  note         = {30-sentence Hinglish eval set (4 categories).
                  Rubric v2.1: three-ASR consensus, Devanagari-normalised CER,
                  ear-only naturalness. 4.70/5.0 mean intelligibility.}
}

If you use IndicF5, also cite the AI4Bharat IndicF5 paper (see the model card at huggingface.co/ai4bharat/IndicF5).

---

License

Component	License
Eval set, scoring scripts, rubric, inference.py	Free for research; cybernovascnn@gmail.com for commercial use
IndicF5 model weights	See ai4bharat/IndicF5 — attribution required, commercial use restricted
F5-TTS architecture	MIT (SWivid/F5-TTS)
IndicXlit	Apache 2.0 (ai4bharat/IndicXlit)

---

For the technical reader

Full evaluation methodology, per-sentence scores, rubric specification, and limitation accounting: EVALUATION_REPORT.md

Architectural details and the why behind each decision: HOW_IT_WORKS.md