FlowyAIPC
  • English

    • ASR Playground

    ASR Playground is the dedicated testing interface for automatic speech recognition models (such as Sherpa-ONNX) in Herdsman. Upload audio files or record from a microphone to quickly evaluate how the model transcribes speech.

    Launching an ASR Model

    1. Open the Models page.

    2. Locate a model tagged ASR or Speech Recognition (e.g., sherpa-onnx Chinese Non-Streaming Paraformer Small).

    3. Click Launch Now on the model card.

    4. Wait for the status to switch to Ready or Running, then click the card to open the Playground.

      Launch ASR model

    Interface Overview

    ASR Playground interface

    Basic Settings and Input

    • Language (left) — choose from the dropdown (e.g., Auto). Selecting a specific language can improve accuracy.
    • Input mode (top tabs):
      • File — process a local audio file (WAV / MP3 / M4A, etc.).
      • Microphone — record audio live for recognition.

    File Transcription Parameters

    • Recommended input — 16 kHz mono PCM WAV for the most reliable results.
    • Compatible formats — FLAC, MP3, M4A, OGG, WebM (requires backend decoder support).
    • How to run — drop the file into the upload area or click Choose Audio, then click Transcribe.

    Reading the Results

    After upload and transcription, the interface updates as follows:

    ASR Playground result

    Transcript Output

    • Character count — shown under the result title (e.g., 77 characters).
    • Text — the full transcript appears in the text box below.

    Runtime Monitor (right sidebar)

    Once transcription completes, performance data appears in the right panel:

    • Latency — total processing time (e.g., 1726 ms).
    • Hardware usage:
      • CPU — utilization during the run (e.g., 7%).
      • GPU — GPU utilization (e.g., 4%).
      • Memory — current memory consumption (e.g., 26.85 GB).

    Run Log (right bottom)

    The log records details for the operation:

    • INFO Transcription complete in 1726ms — confirms successful completion and total elapsed time.
    • INFO Transcription requested: ripple_tts_cloned.wav — confirms the file being processed.

    FAQ

    Q: Why doesn't my MP3 file return a result?

    • Sherpa offline ASR has specific audio format requirements. Although the UI mentions MP3 support, this depends on the backend decoder. Best practice: convert the audio to 16 kHz mono WAV before uploading.

    Q: How can I tell if the model is running fast enough?

    • Check Latency and RTF (real-time factor) on the right. For a 10-second audio clip with a latency of 1726 ms (1.7 seconds), the model is processing roughly 6× faster than real time — excellent performance.