Running the Audio Analytics container - Qualcomm Dragonwing Documentation

Qualcomm’s Audio Analytics container exposes speech recognition (ASR), text translation (T2T), and text-to-speech (TTS) through a REST and WebSocket API. All models run fully on-device on the Qualcomm NPU.

Key features

Three services in one container: ASR, text-to-text translation, and TTS exposed through a single API
NPU-accelerated: Whisper, OpusMT, MeloTTS, and PiperTTS models all run on the Qualcomm NPU, freeing the CPU for other tasks
Multilingual: English, Spanish, and Chinese supported across all three services; German and Italian available via PiperTTS
Streaming support: Real-time transcription via WebSocket alongside batch file-based transcription
OpenAPI / Swagger UI: Interactive API docs available at http://<device-ip>:8085/docs

Supported models

Service	Model	Languages	Sample Rate
ASR	`whisper-small-quantized`	English, Spanish, Chinese	16000 Hz
Translation	`opus-en-es`, `opus-es-en`, `opus-en-zh`, `opus-zh-en`	English, Spanish, Chinese	n/a
TTS	`melo-tts-en`, `melo-tts-es`, `melo-tts-zh`	English, Spanish, Chinese	44100 Hz
TTS	`piper-tts-en`, `piper-tts-de`, `piper-tts-it`	English, German, Italian	22050 Hz

These are the currently supported models. Query /transcriptions/models, /translations/models, and /tts/models to confirm which models are loaded at runtime.

Setup

The steps below assume you have already set up your device (IQ-9075 EVK or IQ-8275 EVK). Once your device is set up, follow these steps to deploy and run the Audio Analytics container.

Install required packages

SSH into your device and install the required packages:

sudo apt update
sudo apt install curl unzip docker-compose

Set up Docker

# Configure Docker group
sudo groupadd docker
sudo usermod -aG docker $USER
newgrp docker

Download the models

Create the model directory structure and download all models from QAI Hub.

IQ-9075 EVK

mkdir -p ~/models/audio/asr ~/models/audio/tts ~/models/audio/translation
cd ~/models

# ASR — Whisper Small
curl -L -o whisper_small_quantized.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/whisper_small_quantized/releases/v0.57.0/whisper_small_quantized-voice_ai-w8a16-qualcomm_qcs9075.zip
unzip -q whisper_small_quantized.zip -d audio/asr && rm whisper_small_quantized.zip

# TTS — MeloTTS (English, Spanish, Chinese)
curl -L -o melotts_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/melotts_en/releases/v0.57.0/melotts_en-voice_ai-mixed_with_float-qualcomm_qcs9075.zip
unzip -q melotts_en.zip -d audio/tts && rm melotts_en.zip

curl -L -o melotts_es.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/melotts_es/releases/v0.57.0/melotts_es-voice_ai-mixed_with_float-qualcomm_qcs9075.zip
unzip -q melotts_es.zip -d audio/tts && rm melotts_es.zip

curl -L -o melotts_zh.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/melotts_zh/releases/v0.57.0/melotts_zh-voice_ai-mixed_with_float-qualcomm_qcs9075.zip
unzip -q melotts_zh.zip -d audio/tts && rm melotts_zh.zip

# TTS — PiperTTS (English, German, Italian)
curl -L -o pipertts_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pipertts_en/releases/v0.57.0/pipertts_en-voice_ai-float-qualcomm_qcs9075.zip
unzip -q pipertts_en.zip -d audio/tts && rm pipertts_en.zip

curl -L -o pipertts_de.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pipertts_de/releases/v0.57.0/pipertts_de-voice_ai-float-qualcomm_qcs9075.zip
unzip -q pipertts_de.zip -d audio/tts && rm pipertts_de.zip

curl -L -o pipertts_it.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pipertts_it/releases/v0.57.0/pipertts_it-voice_ai-float-qualcomm_qcs9075.zip
unzip -q pipertts_it.zip -d audio/tts && rm pipertts_it.zip

# Translation — Opus (English ↔ Spanish, English ↔ Chinese)
curl -L -o opus_mt_en_es.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_en_es/releases/v0.57.0/opus_mt_en_es-voice_ai-float-qualcomm_qcs9075.zip
unzip -q opus_mt_en_es.zip -d audio/translation && rm opus_mt_en_es.zip

curl -L -o opus_mt_es_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_es_en/releases/v0.57.0/opus_mt_es_en-voice_ai-float-qualcomm_qcs9075.zip
unzip -q opus_mt_es_en.zip -d audio/translation && rm opus_mt_es_en.zip

curl -L -o opus_mt_en_zh.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_en_zh/releases/v0.57.0/opus_mt_en_zh-voice_ai-float-qualcomm_qcs9075.zip
unzip -q opus_mt_en_zh.zip -d audio/translation && rm opus_mt_en_zh.zip

curl -L -o opus_mt_zh_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_zh_en/releases/v0.57.0/opus_mt_zh_en-voice_ai-float-qualcomm_qcs9075.zip
unzip -q opus_mt_zh_en.zip -d audio/translation && rm opus_mt_zh_en.zip

IQ-8275 EVK

mkdir -p ~/models/audio/asr ~/models/audio/tts ~/models/audio/translation
cd ~/models

# ASR — Whisper Small
curl -L -o whisper_small_quantized.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/whisper_small_quantized/releases/v0.57.0/whisper_small_quantized-voice_ai-w8a16-qualcomm_qcs8275.zip
unzip -q whisper_small_quantized.zip -d audio/asr && rm whisper_small_quantized.zip

# TTS — MeloTTS (English, Spanish, Chinese)
curl -L -o melotts_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/melotts_en/releases/v0.57.0/melotts_en-voice_ai-mixed_with_float-qualcomm_qcs8275.zip
unzip -q melotts_en.zip -d audio/tts && rm melotts_en.zip

curl -L -o melotts_es.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/melotts_es/releases/v0.57.0/melotts_es-voice_ai-mixed_with_float-qualcomm_qcs8275.zip
unzip -q melotts_es.zip -d audio/tts && rm melotts_es.zip

curl -L -o melotts_zh.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/melotts_zh/releases/v0.57.0/melotts_zh-voice_ai-mixed_with_float-qualcomm_qcs8275.zip
unzip -q melotts_zh.zip -d audio/tts && rm melotts_zh.zip

# TTS — PiperTTS (English, German, Italian)
curl -L -o pipertts_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pipertts_en/releases/v0.57.0/pipertts_en-voice_ai-float-qualcomm_qcs8275.zip
unzip -q pipertts_en.zip -d audio/tts && rm pipertts_en.zip

curl -L -o pipertts_de.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pipertts_de/releases/v0.57.0/pipertts_de-voice_ai-float-qualcomm_qcs8275.zip
unzip -q pipertts_de.zip -d audio/tts && rm pipertts_de.zip

curl -L -o pipertts_it.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/pipertts_it/releases/v0.57.0/pipertts_it-voice_ai-float-qualcomm_qcs8275.zip
unzip -q pipertts_it.zip -d audio/tts && rm pipertts_it.zip

# Translation — Opus (English ↔ Spanish, English ↔ Chinese)
curl -L -o opus_mt_en_es.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_en_es/releases/v0.57.0/opus_mt_en_es-voice_ai-float-qualcomm_qcs8275.zip
unzip -q opus_mt_en_es.zip -d audio/translation && rm opus_mt_en_es.zip

curl -L -o opus_mt_es_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_es_en/releases/v0.57.0/opus_mt_es_en-voice_ai-float-qualcomm_qcs8275.zip
unzip -q opus_mt_es_en.zip -d audio/translation && rm opus_mt_es_en.zip

curl -L -o opus_mt_en_zh.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_en_zh/releases/v0.57.0/opus_mt_en_zh-voice_ai-float-qualcomm_qcs8275.zip
unzip -q opus_mt_en_zh.zip -d audio/translation && rm opus_mt_en_zh.zip

curl -L -o opus_mt_zh_en.zip \
  https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/opus_mt_zh_en/releases/v0.57.0/opus_mt_zh_en-voice_ai-float-qualcomm_qcs8275.zip
unzip -q opus_mt_zh_en.zip -d audio/translation && rm opus_mt_zh_en.zip

Your models directory should now look like this:

~/models/audio/
├── asr/          ← whisper_small_quantized
├── tts/          ← melotts_en, melotts_es, melotts_zh, pipertts_en, pipertts_de, pipertts_it
└── translation/  ← opus_mt_en_es, opus_mt_es_en, opus_mt_en_zh, opus_mt_zh_en

Get the Docker Compose file

Download the Docker Compose file for your device from the CodeLinaro audio-analytics deploy directory.

IQ-9075 EVK

Option A — download directly on the device:

curl -L -o docker-compose-qcs9100.yml \
  https://git.codelinaro.org/clo/iot-solutions-microservices/audio-analytics/-/raw/main/deploy/docker-compose-qcs9100.yml

Option B — download locally and copy to the device:

scp docker-compose-qcs9100.yml ubuntu@<device-ip>:~/

IQ-8275 EVK

Option A — download directly on the device:

curl -L -o docker-compose-qcs8300.yml \
  https://git.codelinaro.org/clo/iot-solutions-microservices/audio-analytics/-/raw/main/deploy/docker-compose-qcs8300.yml

Option B — download locally and copy to the device:

scp docker-compose-qcs8300.yml ubuntu@<device-ip>:~/

The remaining steps use docker-compose-qcs9100.yml. If you are on the IQ-8275 EVK, substitute docker-compose-qcs8300.yml in each command.

Configure the container

Open the Docker Compose file and adjust the following environment variables as needed:

MODELS_DIR: Base path for all models inside the container (default: /mnt/work/models)
SESSION_TIMEOUT_S: Seconds of inactivity before an ASR streaming session is automatically closed (default: 30)
LOG_LEVEL: 10=DEBUG, 20=INFO, 30=WARNING, 40=ERROR (default: 20)

Under volumes:

Model path: By default maps ~/models/audio on the host to /mnt/work/models in the container. If you placed your models elsewhere, update the left side of the volume mount:

volumes:
  - ~/models/audio:/mnt/work/models:ro

Make sure the volume mount points to your models directory and that the directory is readable by the container. If models fail to load, check directory permissions with ls -la ~/models/audio.

Start the container

The Docker Compose file pulls the container image automatically if it isn’t already present.

docker compose -f docker-compose-qcs9100.yml up -d

Verify the container is running:

docker ps | grep audio-analytics

Verify the service is running

Once ready, check the health endpoint:

curl http://localhost:8085/audio-analytics/v1/api/health

Expected response:

{ "status": "ok", "uptime": 12345, "active_sessions": 0, "version": "1.0.0" }

You can also confirm which models are loaded:

curl http://localhost:8085/audio-analytics/v1/api/transcriptions/models
curl http://localhost:8085/audio-analytics/v1/api/translations/models
curl http://localhost:8085/audio-analytics/v1/api/tts/models

Test each service

Run a quick smoke test for each of the three services.ASR — transcribe audio:

curl -X POST http://localhost:8085/audio-analytics/v1/api/transcriptions/create \
  -F 'file=@audio.wav' \
  -F 'model=whisper-small-quantized' \
  -F 'language=en' \
  -F 'stream=false'

Expected response:

{ "text": "Your transcribed text here.", "language": "en", "type": "transcript.text.done" }

Translation — translate text:

curl -X POST http://localhost:8085/audio-analytics/v1/api/translations/translate \
  -H 'Content-Type: application/json' \
  -d '{"text": ["Hello world"], "model": "opus-en-es", "source_language": "en", "target_language": "es"}'

Expected response:

{ "translations": [{ "translated_text": "Hola mundo", "source_language": "en", "target_language": "es" }] }

TTS — synthesize speech:

curl -X POST http://localhost:8085/audio-analytics/v1/api/tts/synthesize \
-H 'Content-Type: application/json' \
-d '{"text": "Hello, how are you today?", "model": "melo-tts-en", "language": "en", "sample_rate": 44100}' \
--output speech.pcm

Stop the container

When you are finished:

docker compose -f docker-compose-qcs9100.yml down

Viewing logs

All container output is available via docker logs. There are no separate log files.

# Stream all logs
docker logs audio-analytics -f -t

To adjust log verbosity, set LOG_LEVEL in your Docker Compose file and restart the container:

Value	Level
`10`	DEBUG
`20`	INFO (default)
`30`	WARNING
`40`	ERROR

Next steps

Speech recognition

File-based and real-time streaming transcription

Text-to-speech

Synthesize natural speech in multiple languages

Language translation

Translate text between languages on-device

​Key features

​Supported models

​Setup

​Viewing logs

​Next steps

Speech recognition

Text-to-speech

Language translation

Key features

Supported models

Setup

Viewing logs

Next steps