TTS Engine Comparison
Compare synthesis times and audio samples from the same test environment below.
Multilingual & English Voice Cloning
Section titled “Multilingual & English Voice Cloning”Benchmark Prompts
Section titled “Benchmark Prompts”- Standard Prompt (used for Chatterbox Standard/Turbo/Nano, MOSS-TTS, CosyVoice 3, VoxCPM2, Qwen3-TTS):
“Pain and pleasure are two sides of the same coin. Go on now… flip it. Either way, I’ll let you feel all of me.”
- Fish Audio S2 Pro Prompt (tested with bracket expression tags):
“Pain and pleasure are two sides of the same coin. [laughs] Go on now… flip it. [whispers] Either way, I’ll let you feel all of me.”
Performance & Audio Comparison
Section titled “Performance & Audio Comparison”Timings report both full generation time (total wall-clock seconds from request to finished audio) and Real-Time Factor (RTF), defined as generation time divided by audio duration:
- RTF < 1.0 (bold): The engine generates speech faster than real time (for example,
0.50× RTFrenders a 10-second clip in 5 seconds). Only these engines could keep up with a live voice call, which TomoriBot does not implement today. - RTF > 1.0: Generation takes longer than the spoken audio. TomoriBot sends each voice message as a complete file, so a higher RTF only means a longer wait.
| Engine | Windows Native(1) (RTX 4070 Ti SUPER) | Linux / WSL2 | macOS (Apple Silicon) | Audio Sample |
|---|---|---|---|---|
| Fish Audio S2 Pro | ~8-10 min(2) (~65× RTF) | Untested | Untested | |
| Chatterbox (Turbo, Default) | ~5.0s (8.7s clip) 0.57× RTF | Untested | Untested | |
| Chatterbox (Nano) | ~3.0s (8.0s clip) 0.38× RTF | Untested | Untested | |
| Chatterbox (Standard) | ~6.0s (7.8s clip) 0.77× RTF | Untested | Untested | |
| MOSS-TTS | ~12.0s (8.8s clip) 1.36× RTF | Untested | Untested | |
| CosyVoice 3 | ~6.0s (13.9s clip) 0.43× RTF | Untested | Untested | |
| VoxCPM2 | ~8.0s (7.4s clip) 1.09× RTF | Untested | Untested | |
| Qwen3-TTS | ~10.0s (9.2s clip) 1.09× RTF | Untested | Untested |
- (1) Test Environment: NVIDIA GeForce RTX 4070 Ti SUPER (16 GB GDDR6X, Ada Lovelace) on Windows 11 (native execution) using a 26.6-second 24 kHz mono reference audio sample with matching verbatim transcript.
- (2) Fish Audio S2 Pro: Windows execution runs in uncompiled eager mode (~65× RTF) due to CUDA kernel launch latency across its 76 layer evaluations per token. Running on Linux or WSL2 with OpenAI Triton compiler fusion (
torch.compile) is recommended to avoid this dispatch stall.
Japanese Voice Cloning
Section titled “Japanese Voice Cloning”Japanese Benchmark Prompt
Section titled “Japanese Benchmark Prompt”「そんな顔して……ほんとは私にやられたいんでしょ？ざぁこざぁこ～♡」
Japanese Performance & Audio Comparison
Section titled “Japanese Performance & Audio Comparison”| Engine | Windows Native(1) (RTX 4070 Ti SUPER) | Linux / WSL2 | macOS (Apple Silicon) | Audio Sample |
|---|---|---|---|---|
| IrodoriTTS | ~4.0s (8.5s clip) 0.47× RTF | Untested | Untested |
- (1) Measured in the same RTX 4070 Ti SUPER Windows 11 test environment.
Compare the Engines
Section titled “Compare the Engines”TomoriBot returns complete WAV files. Model sizes count parameters; GPU memory figures are setup guidance. Timings above are from one Windows machine.
Reference lengths include upstream recommendations and runtime limits. Check each engine guide for details.
| Engine | Model size; 16 GB GPU | Languages | Reference clip | Voice sources and controls | Speed / streaming path | Choose it for |
|---|---|---|---|---|---|---|
| Chatterbox | 350M Turbo (default), 110M Nano, or 500M Standard; yes, Nano can use CPU | English | 10 seconds; longer is silently ignored past the 10 s prompt window | Reference cloning, supported event tags; standard model offers CFG/exaggeration | Fast/small focus; wrapper returns full WAV | Small English clone setup or CPU experiments |
| Qwen3-TTS | 1.7B per mode; yes, models swap | 10, including English/Japanese | From 3 seconds; no documented cap | Clone or text-described VoiceDesign | Quality-focused; upstream streaming, wrapper buffers | General-purpose multilingual clone and Japanese VoiceDesign |
| MOSS-TTS | 4B clone + ~1.7B design, swapped; 16 GB is an unverified trial target; 8B flagship likely no | Clone: 31, including Japanese; design: English/Chinese | Not documented upstream; no runtime cap | Clone or text-described VoiceGenerator; clone language tags | Experimental; Local clone has upstream streaming backend, wrapper buffers | Compare MOSS clone quality or English/Chinese voice design |
| IrodoriTTS | ~0.8B current v4.1 Small; ~3-4 GB VRAM observed in one local run | Japanese only | ~30 seconds; trimmed at the checkpoint’s 120 s cap | Clone or VoiceDesign; emoji style cues | Sampling steps trade quality for speed; wrapper buffers | Small-footprint Japanese voices and emoji-driven delivery |
| Fish S2 Pro | 4B; official BF16 default (~16-18 GB), optional INT8 for 16 GB | 83 claimed upstream | 10-30 seconds; no runtime cap | Reference cloning (requires ref transcript), free-form bracket expression tags | Heavy Dual-AR model; requires Linux/WSL2 with Triton for fast synthesis (~65× RTF on Windows eager mode) | Fine-grained expressive cloning; check research-license terms |
| VoxCPM2 | 2B; ~8 GB BF16 reported upstream | 30 | 5-30 seconds; documented range, no runtime cap | Clone, Voice Design, transcript-assisted Ultimate Cloning, delivery instructions | ~0.30 RTF on upstream RTX 4090; upstream streaming, wrapper buffers | One multilingual model with the broadest voice-source controls |
| CosyVoice 3 | 0.5B core; 16 GB comfortable, download/runtime larger | 9, including Japanese, plus Chinese dialects | 3-30 seconds; longer is trimmed to the first 30 s | Clone, cross-lingual clone, natural-language delivery | Low-latency focus; native text/audio streaming upstream, wrapper buffers | A future streaming candidate with cross-lingual cloning |
Model size and language counts follow the Chatterbox, Qwen3-TTS, MOSS, Irodori, Fish S2 Pro, VoxCPM2, and CosyVoice 3 upstream pages. Check each guide for OS, driver, license, model revision, and memory details. A 16 GB GPU cannot necessarily host a TTS model and a large local LLM simultaneously.
The first Qwen3-TTS auto request includes model loading. MOSS pre-downloads both models during setup and warms the clone model at startup by default, but either auto server still has to load the other model after a mode switch. TomoriBot waits up to TTS_SYNTHESIZE_TIMEOUT_MS (default 240000 ms) for each complete response.