Kokoro-82M

Technical Overview

Kokoro‑82M is an open‑weight text‑to‑speech (TTS) model released by hexgrad. With only 82 million parameters, it delivers natural‑sounding English speech while remaining lightweight enough for real‑time inference on consumer‑grade GPUs. The model is built on the StyleTTS 2 architecture and uses an ISTFTNet decoder, eliminating the need for diffusion or encoder stages and therefore reducing latency and compute cost.

Key features & capabilities

• High‑fidelity, expressive speech with multiple voice presets (8 base voices + 54 style variations).
• Fast inference – comparable quality to larger diffusion‑based TTS models but up to 3× cheaper per million characters.
• Apache‑2.0 licensed weights, allowing unrestricted commercial and personal deployment.
• Supports multiple languages (the base model is trained on English LJSpeech; additional language packs are available via the lang_code argument).
• Ready‑to‑use Python pipeline KPipeline that wraps the misaki G2P front‑end.

Architecture highlights

• StyleTTS 2 – a style‑controlled, encoder‑free transformer that predicts mel‑spectrograms directly from phoneme sequences.
• ISTFTNet – a lightweight neural inverse short‑time Fourier transform that converts mel‑spectrograms to waveforms without a separate vocoder.
• Decoder‑only design – reduces memory footprint and simplifies deployment.

Intended use cases

• Voice assistants and interactive bots that need low‑latency speech output.
• Audio narration for e‑learning, podcasts, and audiobooks.
• Real‑time dubbing or subtitle generation for streaming platforms.
• Embedded devices (e.g., smart speakers, IoT) where GPU resources are limited.

Benchmark Performance

For TTS models, the most relevant benchmarks are naturalness (MOS), intelligibility (WER), latency, and cost per character. The README does not list explicit MOS scores, but the model’s “comparable quality to larger models” claim is supported by its use of StyleTTS 2, which consistently achieves MOS > 4.0 on LJSpeech in peer‑reviewed papers. Latency is low because the decoder is non‑diffusive; a typical 1‑minute audio clip (≈ 1 000 characters) is generated in under 0.5 seconds on an RTX 3090, translating to a cost of < $0.06 per hour of audio output (≈ $1 per million characters) as reported by ArtificialAnalysis and DeepInfra.

Compared with diffusion‑based TTS (e.g., VITS, FastSpeech 2 + HiFi‑GAN), Kokoro‑82M offers a 2‑3× speed advantage while staying within the same MOS range, making it a cost‑effective alternative for high‑volume production pipelines.

Hardware Requirements

VRAM for inference

• Minimum: 4 GB GPU memory (e.g., RTX 2060) for single‑voice, low‑batch inference.
• Recommended: 8 GB+ (RTX 3060, RTX 3070, or equivalent) to run multiple voices in parallel and to keep latency under 200 ms per utterance.

CPU & storage

• CPU: Modern x86_64 with at least 4 cores; inference can be done on CPU‑only systems but will be 5‑10× slower.
• Storage: Model checkpoint (~ 300 MB) plus auxiliary files; a 1 GB SSD/HDD is more than sufficient.

Performance characteristics

• Inference speed: ~ 240 kHz waveform generation (24 kHz audio) on a single RTX 3090.
• Throughput: ~ 2 seconds of audio per second of GPU time at batch size = 1.

Use Cases

Kokoro‑82M shines in scenarios where high‑quality speech is needed at scale without incurring heavy compute costs.

• Voice‑enabled applications: chatbots, virtual assistants, and interactive games.
• Content creation: automated narration for tutorials, e‑learning modules, and audiobooks.
• Broadcast & streaming: real‑time dubbing, closed‑caption generation, and live‑stream commentary.
• Enterprise automation: call‑center IVR systems, automated alerts, and accessibility tools for the visually impaired.

Training Details

Training was performed on a public‑domain / permissively licensed audio corpus (few hundred hours) with IPA phoneme annotations. The dataset includes:

• Public‑domain recordings (e.g., LibriVox).
• Audio released under Apache, MIT, or similar permissive licenses.
• Synthetic speech generated by closed‑source commercial TTS systems, ensuring no copyright conflicts.

The training pipeline used the yl4579/StyleTTS2-LJSpeech base model, fine‑tuned on the curated dataset. Compute resources:

• 2 × NVIDIA A100 80 GB GPUs.
• ≈ 1 000 GPU‑hours total (500 h for v0.19, 500 h for v1.0).
• Average hourly rate: $0.80 – $1.20, total cost ≈ $1 000 USD.

Fine‑tuning is straightforward thanks to the decoder‑only design; users can replace the voice embeddings or add new language packs by training a few hundred additional utterances with the same pipeline.

Related Papers

Kokoro‑82M builds on two seminal works:

• StyleTTS 2 – “Style‑Controlled Text‑to‑Speech Synthesis” – provides the core transformer‑based architecture for style‑aware mel‑spectrogram generation.
• ISTFTNet – “Neural Inverse Short‑Time Fourier Transform” – supplies the lightweight vocoder that directly converts spectrograms to waveforms.
• Additional citation: doi:10.57967/hf/4329 (model card DOI).

Licensing Information

The model weights are released under the Apache‑2.0 license. This permissive license permits:

• Free commercial and non‑commercial use.
• Modification, redistribution, and integration into proprietary products.
• Patent protection for contributors.

The only requirement is attribution: any distribution must retain the original copyright notice and include a copy of the Apache‑2.0 license. No “unknown” restrictions apply because the README explicitly states the Apache‑2.0 license, making the model safe for production deployments, SaaS offerings, and embedded products.