What's New

• Low Precision (fp16) Execution for Voice Isolation Models: reduced CPU execution time by 38–50% for all Voice Isolation models on avx512 and armv8a based hardware when the enableLowPrecisionExecution flag is enabled, allowing approximately 1.6–2× more concurrent audio streams per instance. On armv8a, this extends the optimization introduced in the previous release to all Voice Isolation models. On Linux x64, the optimization is available in the glibc gcc12 and musl libc-based builds. It is not supported in the glibc gcc7-based build. Hardware support is detected automatically at runtime, so callers don't need to check for it themselves on unsupported hardware or models, the flag is simply a no-op and processing continues at native precision. Defaults to false.
• For Linux x64 systems based on GCC 9-11, Node.js SDK loader will automatically pick up gcc-9 based library. These builds do not support avx512.
• Accent Detection API: requires approximately 1 second of audio to produce a result.
• Language Detection API: requires approximately 4 seconds of audio to produce a result.
• Accent Detection API: requires approximately 4 seconds of audio to produce a result.