Inflect-Micro-v2：仅用936万参数实现完整文本转语音

开发者Owen Song独立完成并开源了Inflect-Micro-v2，这是一款参数量低于1000万的完整文本转语音（TTS）合成模型。1该模型包含9,356,513个可部署参数，FP32格式下占用37.53 MB的存储空间。1

Inflect-Micro-v2在性能和易用性上具有多项优势。1模型支持确定性种子设置、长文本处理以及CPU和CUDA两种推理方式，采用英文音素前端和端到端架构设计。1在盲测评估中，该模型获得66.2%的人类偏好率（21胜10负3平），UTMOS22评分达到4.395分。1在自动语音识别（ASR）智能度测试中，经Qwen3-ASR、Nemotron 3.5和Whisper分别检测的错误率为2.52%、5.45%和2.73%。1模型输出规格为24 kHz单声道音频，采用Apache-2.0开源许可证，已发布在Hugging Face平台。1

Owen Song has independently developed and open-sourced Inflect-Micro-v2, a text-to-speech synthesis model featuring fewer than 10 million parameters 1. The model achieves end-to-end voice generation while maintaining a compact footprint of 9,356,513 deployable parameters, consuming 37.53 MB in FP32 format 1. Released under the Apache-2.0 license on Hugging Face, Inflect-Micro-v2 supports deterministic seeding, long-form text processing, and inference across both CPU and CUDA environments 1.

The model produces 24 kHz mono audio output and incorporates an English phoneme frontend architecture 1. In human blind preference testing, Inflect-Micro-v2 achieved a 66.2% preference rate, with 21 wins, 10 losses, and 3 draws compared to alternatives 1. It scored 4.395 on the UTMOS22 evaluation benchmark 1. Speech recognition intelligibility testing showed error rates of 2.52% with Qwen3-ASR, 5.45% with Nemotron 3.5, and 2.73% with Whisper 1.