Text To Speech
Our Text-to-Speech software solution convert text into synthetic speech to enable AI assistant voice output, automate responses and improve user interaction across digital platforms and services.
软件详情
Features
MPU (i.MX 8 and i.MX 9)
- Multi-speakers (language-dependent)
- Frequency sampling: 16 kHz or 22 kHz
- Languages: English, Mandarin Chinese, Hindi, Spanish, French, Arabic, Portuguese, Russian, German, Italian, Dutch, Catalan, Czech, Welsh, Danish, Greek, Farsi, Finnish, Hungarian, Indonesian, Icelandic, Georgian, Kazakh, Luxembourgian, Latvian, Malayalam, Nepali, Norwegian, Polish, Romanian, Slovak, Slovenian, Serbian, Swedish, Swahili, Turkish, Ukrainian, Vietnamese
- Natural speech quality
- The TTS model is based on VITS architecture (Variational Inference with adversarial learning for end-to-end Text-to-Speech paper). Depending on the language, there may be several speakers, male or female, different accents and tones of voice. The frequency sampling rate is 16 kHz or 22 kHz depending on the model
MCU (i.MX RT700)
- Leverages NPU
- Frequency sampling: 16 kHz
- Language: English
支持的器件
- i.MX8M: i.MX 8M系列 - Arm® Cortex®-A53、Cortex-M4、音频、语音、视频
- i.MX91: i.MX 91系列安全高效，为成千上万的边缘应用提供了Linux®的基本功能。
- i.MX93: i.MX 93应用处理器系列——Arm® Cortex®-A55，ML加速，高能效MPU
- i.MX95: i.MX 95应用处理器系列：支持功能安全的高性能平台，带eIQ® Neutron NPU
- i.MX-RT700: i.MX RT700跨界MCU，配备Arm® Cortex®-M33、NPU、DSP和GPU内核
系统需求
- MCU: TensorFlow Lite
- MPU: ONNX
Footprint and Profiling
| Model | Model size [parameters] | Platform / Core | Real Time Factor | DNS-MOS | Quantization format | Code size [MB] | Library dependency |
|---|---|---|---|---|---|---|---|
| MPU | ~20M | i.MX 95 / 6x Cortex-A55 | 0.26 | 4.35 | 8-bit int (linear layers) | ~23 | ONNX |
| - | - | i.MX 8MP / 4x Cortex-A53 | 0.52 | - | - | - | - |
| - | - | i.MX 8MM / 4x Cortex-A53 | 0.59 | - | - | - | - |
| - | - | i.MX 8MN / 4x Cortex-A53 | 0.62 | - | - | - | - |
| - | - | i.MX 91 / 1x Cortex-A55 | 0.97 | - | - | - | - |
| MCU | ~2.2M | Cortex-M33/NPU | 0.24 | 4.17 | 8-bit int + 16a_8w int | 150 kB + 2.6 MB model size (and 1.7 MB + 2.6 MB SRAM) | TFLite |
