TTS and voice adaptation
Compare, adapt and evaluate voices under real corpus, hardware and deployment constraints.
What we currently think
Voice quality alone does not choose a production model. Adaptation support, data, latency, rights and failure recovery remain load-bearing.
Start with your question
- Which model should I test first?
- Should this voice use zero-shot, LoRA or full training?
- How do I hear whether the adaptation really improved?
Start here
Best Open-Source TTS Models for Production in 2026
A practitioner's comparison of VoxCPM, Qwen3-TTS 1.7B, IndexTTS2, and CosyVoice for voice-profile adaptation on a 24GB GPU. Covers finetuning approach, best checkpoint logic, failure modes, and which model to deploy first.
Which TTS Model Should You Use? A Decision Tree (2026)
Open-source TTS model comparison for 2026 across Qwen3-TTS, CosyVoice, F5-TTS, Fish Speech, Kokoro, Supertonic, VoxCPM, IndexTTS2, Chatterbox, and Higgs Audio. Compare fine-tuning, VRAM, inference speed, latency, edge deployment, and voice quality.
Qwen3-TTS LoRA Fine-Tuning - Scale Sweeps, Checkpoints, and Production Defaults
Qwen3-TTS LoRA fine-tuning guide for custom voices. Covers dataset requirements, 24GB VRAM settings, LoRA scale 0.3 to 0.35, LR override 2e-6, sft_12hz.py bugs, and the companion repo.
Continue exploring
Audio8 TTS 0.6B: Mac, RTX 3090 Ti and LoRA Fine-Tuning Test
We tested Audio8 TTS Preview 0.6B on a 16 GB M2 MacBook Pro and an RTX 3090 Ti, then fine-tuned it on 128 NSC-derived voice clips. See inference, memory, LoRA, ASR, speaker similarity, and checkpoint failure results.
Supertonic 3 On-Device TTS Reality Check on macOS
A first-party Supertonic 3 smoke test on Apple Silicon macOS: ONNX Runtime CPU setup, model cache size, RAM usage, audio output, no-discrete-VRAM behavior, and where it fits against Qwen3-TTS, Kokoro, F5-TTS, and CosyVoice.
LoRA vs Full SFT for Voice Models - What Actually Changes on a 24 GB GPU
A practical guide to choosing LoRA or full supervised fine-tuning for TTS and voice cloning models on RTX 3090 Ti and RTX 4090 class GPUs, using first-party Qwen3-TTS, VoxCPM, CosyVoice, and IndexTTS2 runs.
VoxCPM 2 Full SFT on a 24 GB GPU - The Run That Actually Fit
First-party notes from a VoxCPM 2 full supervised fine-tuning run on an RTX 3090 Ti, including the memory failures, the working PagedAdamW8bit recipe, dataset cleanup, validation results, and checkpoint selection.
How Open-Source TTS Architectures Differ - And What It Means for Fine-Tuning (2026)
A practitioner's comparison of 6 TTS architectures (Voxtral, Qwen3-TTS, IndexTTS2, Chatterbox, Fish Speech, CosyVoice 3) covering codecs, LoRA compatibility, data pipelines, and license traps. Based on first-party adaptation work on the same FEMALE_01 corpus slice.
F5-TTS Fine-Tuning Guide 2026 - Colab, Quality, VRAM, and Voice Cloning
F5-TTS fine-tuning guide for custom voice cloning in 2026. Covers Colab-style setup, dataset preparation, VRAM, quality review, ease of fine-tuning, and comparison with Qwen3-TTS, VoxCPM, and CosyVoice.
CosyVoice Fine-Tuning Guide - LoRA, Data Requirements, and Voice Quality
CosyVoice fine-tuning guide for voice cloning on a consumer GPU. Covers LoRA vs full SFT, data requirements, VRAM, voice quality, the epoch 12 rerun, 9 known pitfalls, and the PEFT companion repo.
Voice Cloning on a 24GB GPU - What Actually Works in 2026
Real VRAM numbers, training times, and recipe availability for VoxCPM, Qwen3-TTS, IndexTTS2, and CosyVoice on an RTX 3090 Ti. Which models fit, which need LoRA vs full SFT, and the failure modes to expect.
Designing a Contract-First TTS Layer for Production Video Pipelines
Most teams evaluate TTS by demo quality alone. A production narration stack needs runtime contracts, cue timelines, artifact sidecars, and QA references so the workflow survives engine changes and timing drift.
MOSS-TTS First Technical Read and Production Reality Check
A first engineering read of OpenMOSS MOSS-TTS as of February 2026: what is actually released, what benchmark claims are reported, where deployment risk is still high, and how we will follow up after a 24GB feasibility smoke test.
GLM-TTS Technical Report for Production Zero-Shot TTS
A practical read of GLM-TTS (arXiv:2512.14291): what is actually novel, what the open benchmark numbers show, what is still internal-only, and how to evaluate fit for production voice workflows.
ReStyle-TTS and Relative Style Control in Zero-Shot TTS
A practical read on ReStyle-TTS (arXiv:2601.03632): what is novel, what the reported results show, what this unlocks for voice workflows, and what still blocks production adoption until code or demo artifacts are published.
CosyVoice 2 vs 3 - Voice Cloning Quality Compared (2026)
CosyVoice 2 vs CosyVoice 3 voice cloning quality review for 2026. Includes audio samples, zero-shot baseline notes, fine-tuning results, LoRA rerun status, and when to use each model.
IMDA NSC Voice Cloning Finetuning Benchmark 2026
A practical benchmark of CosyVoice, VoxCPM, Qwen3-TTS, and IndexTTS2 fine-tuned on IMDA NSC FEMALE_01, with operational notes for both product teams and engineers.
IndexTTS2 Finetuning on IMDA NSC FEMALE_01
Practical notes from our IndexTTS2 single-speaker finetuning run on IMDA NSC FEMALE_01, including crash recovery, checkpoint retention behavior, and checkpoint selection.
VoxCPM 1.5 LoRA Finetuning on IMDA NSC FEMALE_01
Exact run notes for VoxCPM 1.5 LoRA finetuning on IMDA NSC FEMALE_01, including checkpoint selection, prompt/no-prompt behavior, and denoiser tradeoffs.
Voice Cloning Finetuning Guide: E2-TTS, F5-TTS, and GPT-SoVITS V2Pro
SpeechRole (Aug 2025) puts E2-TTS and F5-TTS at the top for controllable voice cloning when you have speaker data. Here's how to compare them with the latest GPT-SoVITS V2Pro release and pick the right stack for production.
CosyVoice 3 Explained - Architecture, Training, and What to Expect
How CosyVoice 3 works from zero - speech tokens, flow matching, DiffRO reward optimisation, and practical deployment notes. Written for engineers evaluating open-source TTS, not just reading the paper.