fix: Use explicit Int64 tensor type for Silero VAD sample rate input

[EgilSandfeld]

The Silero VAD ONNX model expects tensor(int64) for the 'sr' input, but was receiving tensor(float) due to implicit type inference issues.

• Use OrtValue.CreateTensorValueFromMemory() to explicitly create Int64 tensor
• Add using directive for Microsoft.ML.OnnxRuntime.Tensors
• Fixes ONNX Runtime input type mismatch error

Problem

The Silero VAD ONNX model expects tensor(int64) for the 'sr' (sample rate) input, but was receiving tensor(float) due to implicit type inference issues in ONNX Runtime. This caused the following error:

Unexpected input data type. Actual: (tensor(float)) , expected: (tensor(int64))

Solution

Use OrtValue.CreateTensorValueFromMemory<long>() to explicitly create an Int64 tensor for the sample rate input.

Changes

• Use OrtValue.CreateTensorValueFromMemory<long>(sampleRateInput, [1]) to explicitly create Int64 tensor
• Add using directive for Microsoft.ML.OnnxRuntime.Tensors

Testing

• Build succeeds on all target frameworks (netstandard2.0, net8.0, net9.0)
• All existing tests pass (99 passed on net9.0, 101 passed on net472)

[sandrohanea]

Hello @EgilSandfeld and thanks for the contribution!
The fix is indeed great.

However, the standardize target framework to netstandard2.0 is breaking existing optimizations that are using preprocessor directives:

e.g.

#if NET8_0_OR_GREATER
    [LibraryImport("sherpa-onnx-c-api", EntryPoint = "SherpaOnnxAcceptWaveformOffline")]
    private static partial void AcceptWaveform(IntPtr handle, int sampleRate, IntPtr samples, int n);
#else
    [DllImport("sherpa-onnx-c-api", EntryPoint = "SherpaOnnxAcceptWaveformOffline")]
    private static extern void AcceptWaveform(IntPtr handle, int sampleRate, IntPtr samples, int n);
#endif

Place:

echosharp/components/EchoSharp.Onnx.Sherpa/SpeechTranscription/SherpaOnnxOfflineSpeechTranscriptor.cs

Line 16 in 2a7997d

#if NET8_0_OR_GREATER

`
During build, now all code will use the else (DllImport code).

This is just some small example without much perf improvements, but there are other optimizations that really make more sense:

echosharp/src/EchoSharp/Audio/SilenceAudioSource.cs

Line 101 in 2a7997d

#if NET8_0_OR_GREATER

[sandrohanea]

Why adding these 2 different versions for the same package?

[sandrohanea]

this should be a transitive dependency from Whisper.net already for the required frameworks.

[sandrohanea]

Why is this needed on dotnet target frameworks (net8 / net9) as a dependency?
I think those already have support for the required IAsyncEnumerable and do not require additional dependency.