Streaming Speech Translation Pipeline

Real-time English → Russian speech translation: Audio In → ASR → NMT → TTS → Audio Out Translates spoken English into spoken Russian with streaming output over WebSocket.

Input can only be English for now (due to ASR NeMo), while output language depending on TranslateGemma (NMT) and Qwen3-TTS/XTTSv2 (TTS). You can modify these accordingly.

Updates

• As of 2026/05/05, you can use Qwen3-TTS, which is based on our Qwen3-TTS-Streaming-ONNX. This is integrated with the modular TTS implementation, where you can plug-in your favorite streaming TTS module.
• As of 2026/05/14, the NMT has been made modular, so, you can plug-in your favorite NMT module for streaming, similarly as TTS.

Architecture

Audio Input → ASR (ONNX)                           → NMT (GGUF/vLLM)        → TTS (ONNX)       → Audio Output
(PCM16)       NeMo-Cache-Aware-FastConformer-RNN-T   TranslateGemma/Gemma-4   Qwen3-TTS/XTTSv2   (PCM16)

• ASR: NVIDIA NeMo FastConformer RNN-T (cache-aware streaming, ONNX)
• NMT: TranslateGemma 4B (GGUF Q8_0, llama-cpp-python) with streaming segmentation and translation merging or Gemma-4-E4B-it (vLLM with MTP)
• TTS: Qwen3-TTS-Streaming or XTTSv2-Streaming (ONNX), 24kHz output

See ARCHITECTURE.md for detailed design documentation.

Requirements

• Python 3.12
• CUDA 12.8, CUDNN 9
• llama-cpp-python with GPU
• onnxruntime with GPU
• Model files:
  • ASR: NeMo Cache-Aware FastConformer RNN-T ONNX model directory
  • NMT: TranslateGemma 4B GGUF file or vLLM server for Gemma-4-E4B-it with MTP
  • TTS: Qwen3-TTS/XTTSv2 ONNX model and configs directory

Installation

• For CUDA and CUDNN:

# uninstall, remove, and purge previous CUDA and CUDNN installation if they are not the correct versions
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
sudo apt-get install -y cuda-toolkit-12-8 cudnn9-cuda-12
echo 'export LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}' >> ~/.bashrc
echo 'export PATH=/usr/local/cuda-12.8/bin:$PATH' >> ~/.bashrc
source ~/.bashrc

• For llama-cpp-python (GGUF):

echo 'export CMAKE_ARGS="-DGGML_CUDA=on"' >> ~/.bashrc
echo 'export CUDACXX=/usr/local/cuda/bin/nvcc' >> ~/.bashrc
echo 'export FORCE_CMAKE=1' >> ~/.bashrc
source ~/.bashrc
sudo apt-get install -y build-essential cmake ninja-build python3-dev
pip uninstall -y llama-cpp-python
pip install --no-cache-dir --force-reinstall llama-cpp-python

• For this model package:

pip install -r requirements.txt

System Dependencies

# Ubuntu/Debian
apt-get install libsndfile1 libportaudio2

Usage

Start the Server

• Recommended to use at least g6.xlarge (L4, 4vCPUs, 16GB RAM). GPU memory usage will be around 17GB/9GB for all 3 ASR-NMT-Qwen3TTS/XTTSv2 models.
• Total model latencies will be around 350/300 ms (ASR 18 ms., NMT 150 ms., Qwen3-TTS/XTTSv2 200/150 ms.). g5.xlarge with A10G will have a 1.5x faster processing time.
• This, still bounded by the ASR lookahead 560 ms (NeMO Cache-Aware Streaming ASR FastConformer-RNN-T).
• Hence, the effective delay is 560+350/300=910/860 ms, which is very comfortable for a simultaneous/streaming speech translation.

python app.py \
  --asr-onnx-path models/asr/nemo-cache-aware-streaming-560ms-onnx/ \
  --streaming-nmt-config-path configs/nmt/translategemma/translategemma_streaming_nmt_config.json \
  --streaming-nmt-pipeline-config-path configs/nmt/translategemma/translategemma_streaming_nmt_pipeline_config.json \
  --streaming-tts-config-path configs/tts/qwen3_tts/qwen3_tts_streaming_tts_config.json \
  --streaming-tts-pipeline-config-path configs/tts/qwen3_tts/qwen3_tts_streaming_tts_pipeline_config.json \
  --tts-ref-audio-path audio_ref/male_stewie.mp3 \
  --host 0.0.0.0 \
  --port 8765

CLI Options

• You can find TTS model specific related configurations in the configs/tts/qwen3_tts/ or configs/tts/xtts/ dir., which are used for the streaming-tts-config-path and streaming-tts-pipeline-config-path.
• These include TTS's model location, queue sizes, chunk sizes, number of threads, and gpu usage.

Python Client

Captures microphone audio and plays back translated speech:

pip install -r requirements_client.txt
python clients/python_client.py --uri ws://<server_ip_address/localhost>:8765

Web Client

TBD

WebSocket Protocol

Docker

TBD

Project Structure

streaming_speech_translation/
├── app.py                              # Main entry point
├── requirements.txt
├── README.md
├── ARCHITECTURE.md
├── Dockerfile
├── models/
│   ├── asr/
│   │   └── nemo-cache-aware-streaming-560ms-onnx/
│   ├── nmt/
│   │   ├── translategemma-4b-it-q8_0-gguf/
│   │   └── translategemma-4b-it-q4_k_m-gguf/
│   └── tts/
│       ├── qwen3-tts-onnx/
│       └── xttsv2-onnx/
├── src/
│   ├── asr/
│   │   ├── streaming_asr.py            # StreamingASR wrapper
│   │   ├── cache_aware_modules.py      # Audio buffer + streaming ASR
│   │   ├── cache_aware_modules_config.py
│   │   ├── modules.py                  # ONNX model loading
│   │   ├── modules_config.py
│   │   ├── onnx_utils.py
│   │   └── utils.py                    # Audio utilities
│   ├── nmt/
│   │   ├── base
│   │   │   ├── streaming_nmt_base.py               # Base class for StreamingNMT wrapper
│   │   │   └── streaming_nmt_pipeline_base.py      # Base class for StreamingNMTPipeline wrapper
│   │   ├── factory
│   │   │   ├── streaming_nmt_factory.py            # Factory class for StreamingNMT wrapper
│   │   │   └── streaming_nmt_pipeline_factory.py   # Factory class for StreamingNMTPipeline wrapper
│   │   ├── pipeline
│   │   │   └── translategemma/
│   │   │        └── translategemma_nmt_streaming_pipeline.py # TranslateGemma-specific StreamingNMTPipeline class
│   │   ├── streaming
│   │   │   └── translategemma/
│   │   │        └── translategemma_streaming_nmt.py          # TranslateGemma-specific StreamingNMT class
│   │   └── utils/
│   │       ├── text_utils.py                    
│   │       └── translategemma/
│   │            ├── streaming_segmenter.py             # ASR output segmenter for NMT input
│   │            └── streaming_translation_merger.py    # Handler to stabilize NMT output revision
│   ├── tts/
│   │   ├── base
│   │   │   ├── streaming_tts_base.py               # Base class for StreamingTTS wrapper
│   │   │   └── streaming_tts_pipeline_base.py      # Base class for StreamingTTSPipeline wrapper
│   │   ├── factory
│   │   │   ├── streaming_tts_factory.py            # Factory class for StreamingTTS wrapper
│   │   │   └── streaming_tts_pipeline_factory.py   # Factory class for StreamingTTSPipeline wrapper
│   │   ├── pipeline
│   │   │   ├── qwen3_tts/
│   │   │   │    └── qwen3_tts_streaming_pipeline.py# Qwen3-TTS-specific StreamingTTSPipeline class
│   │   │   └── xtts/
│   │   │        └── xtts_streaming_pipeline.py     # XTTSv2-specific StreamingTTSPipeline class
│   │   ├── streaming
│   │   │   ├── qwen3_tts/
│   │   │   │    └── qwen3_tts_streaming_tts.py     # Qwen3-TTS-specific StreamingTTS class
│   │   │   └── xtts/
│   │   │        └── xtts_streaming_tts.py          # XTTSv2-specific StreamingTTS class
│   │   └── utils/
│   │       ├── tts_segmenter.py                    # Handler for TTS text input segmenter
│   │       ├── qwen3_tts/
│   │       │    ├── audio_utils.py                 # Audio utility functions for Qwen3-TTS
│   │       │    ├── qwen3_tts_config.py            # Model and sampling configs dataclass for Qwen3-TTS
│   │       │    ├── qwen3_tts_onnx_session_manager.py # ONNX sessions handler for Qwen3-TTS modules
│   │       │    ├── qwen3_tts_text_processor.py    # Qwen3-TTS-specific text front-end
│   │       │    └── utils.py                       # Qwen3-TTS-specific utilities functions (sampling, RoPE)
│   │       └── xtts/
│   │            ├── xtts_onnx_orchestrator.py      # ONNX sessions handler for XTTSv2 modules
│   │            ├── xtts_tokenizer.py              # XTTSv2 BPE tokenizer
│   │            ├── xtts_tts_warmup.py             # Warmup handler for XTTSv2 ONNX sessions
│   │            └── zh_num2words.py                # XTTSv2 Chinese text normalization
│   ├── pipeline/
│   │   ├── orchestrator.py             # PipelineOrchestrator
│   │   └── config.py                   # PipelineConfig
│   └── server/
│       └── websocket_server.py         # WebSocket server
└── clients/
    ├── python_client.py                # Python CLI client
    └── web_client.html                 # Browser client

See ARCHITECTURE.md for the full concurrency diagram and queue map.

LICENSE and COPYRIGHT

This repository is released under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). This means:

• ✅ Research and academic use
• ✅ Personal experimentation
• ✅ Open-source contributions
• ❌ Commercial applications
• ❌ Production deployment
• ❌ Monetized services

By: Patrick Lumbantobing

Copyright@VertoX-AI

Citation

If you use this system in your research, please cite:

@misc{vertoxai2026streamingspeechtranslation,
  title={Streaming Speech Translation — VertoX-AI},
  author={Tobing, P. L., VertoX-AI},
  year={2026},
  publisher={HuggingFace},
}

Acknowledgments

• NVIDIA for Cache-Aware ASR NeMo
• istupakov for the ONNX reference
• Google for the TranslateGemma NMT model and Gemma-4
• Coqui for the XTTSv2
• Qwen for the Qwen3-TTS