Murivo-TTS 21M English (48 kHz Flow-Matching DiT)

Lightweight Text-to-Speech Engine for English

Murivo-TTS is an ultra-lightweight, 100% offline text-to-speech engine. It combines a ~20.9M parameter Flow-Matching Diffusion Transformer (DiT) acoustic backbone with the VoxCPM2 48 kHz Causal AudioVAE neural vocoder to synthesize 48,000 Hz audio natively.

Model File: murivo_21m_english.pt (~41.9 MB, FP16)
Sampling Rate: 48,000 Hz (VoxCPM2 AudioVAE)
Architecture: Flow-Matching Diffusion Transformer (Euler ODE Solver)
Language: English (US Studio)
Checkpoint Status: Studio Foundation (Step 12,500)

🔬 Training Pipeline & Alignment Methodology

Murivo-TTS English was trained on the single-speaker LJSpeech corpus using the Rápido DiT flow-matching architecture:

G2P Frontend: Uses an offline CMU IPA phonetic dictionary with acronym expansion (AI, GPU, LLM) rather than external espeak-ng runtime dependencies.
Warmup & Freezing: Phase 1 (1,000 steps) froze the DiT backbone to anchor character-to-duration alignment; Phase 2 (11,500 steps) unfroze all blocks for full joint flow-matching training.

🎧 Audio Showcase & Evaluation

Listen to native 48 kHz master synthesis samples generated directly by this checkpoint:

Sample	Audio Player	Transcript & English Translation
Sample 1		Native: Every day gives us another opportunity to learn something new and become better at what we do. English: "Every day gives us another opportunity to learn something new and become better at what we do." Note: Evaluation of continuous multi-clause cadence and natural vowel articulation.
Sample 2		Native: Education opens doors to new opportunities and helps people build a better future. English: "Education opens doors to new opportunities and helps people build a better future." Note: Testing plosive articulation ('opens', 'people') and fluent transitions.
Sample 3		Native: When people work together and share their knowledge, they can achieve great things. English: "When people work together and share their knowledge, they can achieve great things." Note: Evaluating stress timing and phonetic balance.

📊 Training Curriculum & Progression

Dataset: keithito/lj_speech
Curriculum Summary: Trained from scratch for 12,500 steps on LJSpeech (24 hours of single-speaker studio audio). Features a two-phase curriculum: Phase 1 (1,000 steps) frozen-backbone warmup to lock phoneme duration and text alignment, followed by Phase 2 (11,500 steps) full joint harmony flow-matching.

Convergence & Loss Progression

Training Phase	Step Range	Total Loss	Flow Loss	Duration Loss
Phase 1 (Warmup)	Step 1	`1.6732`	`1.4299`	`0.4866`
Phase 1 (Warmup)	Step 1,000	`1.0730`	`1.0305`	`0.0239` (-95.1%)
Phase 2 (Joint)	Step 5,000	`0.9412`	`0.9234`	`0.0078`
Phase 2 (Joint)	Step 10,000	`0.9125`	`0.8792`	`0.0057`
Phase 2 (Final)	Step 12,500	`0.8931`	`0.8744`	`0.0058` (-98.8%)

⚡ Quickstart Guide

1. Requirements

Install the minimal dependencies (standard Python, no C-libraries or system espeak-ng required):

pip install torch soundfile numpy

2. Command-Line Inference (CLI)

Download this repository or infer.py and run:

python infer.py --lang english --text "Every day gives us another opportunity to learn something new and become better at what we do." --out output_english.wav

3. Python API

from infer import MurivoTTS

# Initialize engine (automatically manages model and vocoder)
tts = MurivoTTS(lang="english", device="cuda")  # or device="cpu"

# Synthesize 48 kHz audio waveform
waveform = tts.synthesize(
    text="Every day gives us another opportunity to learn something new and become better at what we do.",
    guidance=1.8,
    speed=1.0,
    steps=20
)

# Save to disk
tts.save_wav(waveform, "output.wav", sample_rate=48000)

Note on Vocoder: The engine decodes speech using the VoxCPM2 48 kHz causal AudioVAE. On first run, infer.py will automatically cache audiovae.pth from Hugging Face (openbmb/VoxCPM2) if it is not already in your directory.

📜 Attributions & Acknowledgments

Trained & Created by: Samuel Olubukun
Acoustic Architecture Lineage: Special thanks to Anurag Kanade and the creators of Rápido TTS for their pioneering lightweight DiT flow-matching speech synthesis architecture. Check out Rápido TTS on Hugging Face and GitHub.
Neural Vocoder: Built upon the 48 kHz causal VoxCPM2 AudioVAE developed by the OpenBMB Team.
Dataset Communities:
- Special thanks to the LJ Speech Dataset by Keith Ito.

📖 Citations

Citing Murivo-TTS

If you use Murivo-TTS in your research or applications, please cite:

@misc{murivo2026,
  title        = {Murivo-TTS: High-Fidelity 48 kHz Flow-Matching Text-to-Speech for English, Yorùbá, Hausa & Igbo},
  author       = {Olubukun, Samuel},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/samuelolubukun/murivo-tts-21m-english}}
}

Architectural Lineage (Rápido TTS)

Murivo's acoustic backbone builds upon the open-source Rápido DiT architecture by Anurag Kanade:

@misc{rapido2026,
  title        = {Rápido: Tiny, Fast Brazilian Portuguese Text-to-Speech},
  author       = {Kanade, Anurag},
  year         = {2026},
  howpublished = {\url{https://github.com/anurag12-webster/rapido-tts}}
}

📄 License

This model and its accompanying code are released under the Apache 2.0 License.