STT Failure: "No Voice detected" when using Wyoming Whisper with Sherpa-ONNX

[pvossel]

Spoken commands fail with the error "No Voice detected" (and Home Assistant logs show stt-no-text-recognized) when the Wyoming STT engine is configured to use the Sherpa-ONNX library (e.g., Parakeet-TDT models).

The wake word is detected perfectly, and the VAD (Voice Activity Detection) triggers correctly in the logs, but no text is transcribed.

Technical Details:
App Version: Ava v0.4.5 (knoop7 fork)
STT Engine: rhasspy/wyoming-whisper:latest
Library: sherpa (specifically sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8)

Behavior:
Switching back to standard faster-whisper library with medium-int8 model resolves the issue immediately.

The brownard/Ava fork works correctly with the Sherpa library, suggesting an audio encoding or streaming protocol mismatch in the knoop7 audio pipeline when communicating with Sherpa-based Wyoming servers.

- type: stt-vad-start
  timestamp: "2026-04-08T07:45:41.956267+00:00"
- type: stt-vad-end
  timestamp: "2026-04-08T07:45:42.835689+00:00"
- type: error
  data:
    code: stt-no-text-recognized
    message: No text recognized

It seems the audio stream format or chunking provided by this fork is not compatible with the stricter requirements of the Sherpa library.

[knoop7]

https://github.com/knoop7/Ava/releases/download/0.4.6/Ava-0.4.6-release.apk

[pvossel]

Sadly 0.4.6 did not resolve the issue with "sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8".
When I change Wyoming Whisper to model "medium-int8" it works.

sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8 runs way better and faster on my CPU-only system.

[knoop7]

Sadly 0.4.6 did not resolve the issue with "sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8". When I change Wyoming Whisper to model "medium-int8" it works.

sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8 runs way better and faster on my CPU-only system.

Sorry, I continue to try to solve this problem, I will not give up, and I am still solving it in small ways, gradually implementing more voice customization, and solving your problem

Could you please show me how to install the STT method in HASS? How did you install it? Did you use any third-party containers or integrations?

[pvossel]

No need to apologize! Your fork is great because of the ESPHome native API integration, which I really value for my tablet sensors.

I am running Wyoming Whisper standalone in a K3S (Kubernetes) cluster using the official image: rhasspy/wyoming-whisper:latest.

The Issue
AVA (knoop7): The pipeline breaks at the STT stage. Logs show stt-no-text-recognized. It seems the audio stream from this fork isn't compatible with the sherpa library's requirements.

AVA (Brownard): Works perfectly with the exact same Sherpa/Parakeet setup.

Workaround: If I switch the library back to faster-whisper with the medium-int8 model, your fork works again.

How to reproduce (STT Setup)
In Home Assistant, I use the official Wyoming Protocol integration pointing to my K3S node IP and port.

1. Download the model:

wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2
tar xvf sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2

2. Kubernetes Deployment (YAML):
   This configuration forces the use of the sherpa library instead of the default faster-whisper.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: wyoming-whisper
spec:
  replicas: 1
  template:
    spec:
      containers:
        - name: whisper
          image: rhasspy/wyoming-whisper:latest
          args:
            - "--stt-library"
            - "sherpa"
            - "--model"
            - "/data/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8"
            - "--language"
            - "nl"
            - "--cpu-threads"
            - "5"
            - "--data-dir"
            - "/data"
          ports:
            - containerPort: 10300

Model file structure for reference:

-rw-r--r-- decoder.int8.onnx (12M)
-rw-r--r-- encoder.int8.onnx (622M)
-rw-r--r-- joiner.int8.onnx  (6.1M)
-rw-r--r-- tokens.txt        (92K)

I compared your MicrophoneInput.kt with Brownard's. I found the difference: You are using AcousticEchoCanceler, AutomaticGainControl, and NoiseSuppressor in setupAudioEffects().

While these effects are usually helpful, they often distort the raw PCM data in a way that modern, sensitive models like Sherpa-ONNX / Parakeet cannot handle.

Specifically:

NoiseSuppressor (NS) often filters out frequencies that the model needs.

AutomaticGainControl (AGC) messes with the volume levels, causing the model to lose track.

Brownard does not use any of these effects, sending the 'pure' raw stream instead. My guess is that the stt-no-text-recognized error happens because the audio is being 'over-processed' by Android before it reaches the Wyoming server.

Could you try a build where these effects are opt-in opt-out to see if that fixes the Sherpa compatibility?