collabora/WhisperLive

This project is a real-time transcription application that uses the OpenAI Whisper model to convert speech input into text output. It can be used to transcribe both live audio input from microphone and pre-recorded audio files.

• Installation
• Getting Started
• Running the Server
• Running the Client
• Advanced Features
  • Word-Level Timestamps
  • Custom Vocabulary / Hotwords
  • Speaker Diarization
  • Batch Inference
  • Raw PCM Input
• Streaming Client (Manual Audio Chunking)
• Browser Extensions
• Aavaaz
• Whisper Live Server in Docker
• Troubleshooting
• Future Work
• Blog Posts
• Contact
• Citations

• Install PortAudio (required system dependency for microphone input via PyAudio)

 bash scripts/setup.sh

On Debian/Ubuntu this installs portaudio19-dev, on Fedora portaudio-devel, on macOS it uses Homebrew (portaudio).

• Install 3.12 venv (on Fedora sudo dnf install -y python3.12 python3.12-pip)

python3.12 -m venv whisper_env
source whisper_env/bin/activate

• Install whisper-live from pip

 pip install whisper-live

python3 run_server.py --port 9090 --backend faster_whisper --max_clients 4 --max_connection_time 600 --enable_rest --cors-origins="http://localhost:8080,http://127.0.0.1:8080"

python3 client_openai.py $AUDIO_FILE

• Please follow TensorRT_whisper readme for setup of NVIDIA/TensorRT-LLM and for building Whisper-TensorRT engine.

The server supports 3 backends faster_whisper, tensorrt and openvino. If running tensorrt backend follow TensorRT_whisper readme

• Faster Whisper backend

python3 run_server.py --port 9090 \
                      --backend faster_whisper \
                      --max_clients 4 \
                      --max_connection_time 600
  
# running with custom model and cache_dir to save auto-converted ctranslate2 models
python3 run_server.py --port 9090 \
                      --backend faster_whisper \
                      --max_clients 4 \
                      --max_connection_time 600 \
                      -fw "/path/to/custom/faster/whisper/model" \
                      -c ~/.cache/whisper-live/

• TensorRT backend. Currently, we recommend to only use the docker setup for TensorRT. Follow TensorRT_whisper readme which works as expected. Make sure to build your TensorRT Engines before running the server with TensorRT backend.

# Run English only model
python3 run_server.py -p 9090 \
                      -b tensorrt \
                      -trt /home/TensorRT-LLM/examples/whisper/whisper_small_en \
                      --max_clients 4 \
                      --max_connection_time 600

# Run Multilingual model
python3 run_server.py -p 9090 \
                      -b tensorrt \
                      -trt /home/TensorRT-LLM/examples/whisper/whisper_small \
                      -m \
                      --max_clients 4 \
                      --max_connection_time 600

Note: The TensorRT backend uses a C++ session by default. If you experience issues (e.g. repeated CrossAttentionMask warnings or crashes), add the --trt_py_session flag to use the Python session instead.

• Use --max_clients option to restrict the number of clients the server should allow. Defaults to 4.
• Use --max_connection_time options to limit connection time for a client in seconds. Defaults to 600.
• WhisperLive now supports the OpenVINO backend for efficient inference on Intel CPUs, iGPU and dGPUs. Currently, we tested the models uploaded to huggingface by OpenVINO.

  • Docker Recommended: Running WhisperLive with OpenVINO inside Docker automatically enables GPU support (iGPU/dGPU) without requiring additional host setup.

  • Native (non-Docker) Use: If you prefer running outside Docker, ensure the Intel drivers and OpenVINO runtime are installed and properly configured on your system. Refer to the documentation for installing OpenVINO.

python3 run_server.py -p 9090 -b openvino

• Please follow ROCm_whisper readme for setup of AMD ROCm GPU support with the CTranslate2 ROCm wheel.

To control the number of threads used by OpenMP, you can set the OMP_NUM_THREADS environment variable. This is useful for managing CPU resources and ensuring consistent performance. If not specified, OMP_NUM_THREADS is set to 1 by default. You can change this by using the --omp_num_threads argument:

python3 run_server.py --port 9090 \
                      --backend faster_whisper \
                      --omp_num_threads 4

By default, when running the server without specifying a model, the server will instantiate a new whisper model for every client connection. This has the advantage, that the server can use different model sizes, based on the client's requested model size. On the other hand, it also means you have to wait for the model to be loaded upon client connection and you will have increased (V)RAM usage.

When serving a custom TensorRT model using the -trt or a custom faster_whisper model using the -fw option, the server will instead only instantiate the custom model once and then reuse it for all client connections.

If you don't want this, set --no_single_model.

Use the below command to run the client:

python3 run_client.py --files <audio-file-name>

This will connect to the localhost server running on port 9090 by default. Use flags --server and --port to use different configurations. The above command will transcribe audio file provided with --files flag.

Here are the details of client instance implemented in run_client.py script:

• lang: Language of the input audio, applicable only if using a multilingual model.
• translate: If set to True then translate from any language to en.
• model: Whisper model size.
• use_vad: Whether to use Voice Activity Detection on the server.
• save_output_recording: Set to True to save the microphone input as a .wav file during live transcription. This option is helpful for recording sessions for later playback or analysis. Defaults to False.
• output_recording_filename: Specifies the .wav file path where the microphone input will be saved if save_output_recording is set to True.
• mute_audio_playback: Whether to mute audio playback when transcribing an audio file. Defaults to False.
• enable_translation: Start translation thread on the server (from any to any).
• target_language: Server translation thread's target translation language.

from whisper_live.client import TranscriptionClient
client = TranscriptionClient(
  "localhost",
  9090,
  lang="en",
  translate=False,
  model="small",                                      # also support hf_model => `Systran/faster-whisper-small`
  use_vad=False,
  save_output_recording=True,                         # Only used for microphone input, False by Default
  output_recording_filename="./output_recording.wav", # Only used for microphone input
  mute_audio_playback=False,                          # Only used for file input, False by Default
  enable_translation=True,
  target_language="hi",
  initial_prompt=None,                                # Add context for the model, e.g. 'Jane Doe context'
)

It connects to the server running on localhost at port 9090. Using a multilingual model, language for the transcription will be automatically detected. You can also use the language option to specify the target language for the transcription, in this case, English ("en"). The translate option should be set to True if we want to translate from the source language to English and False if we want to transcribe in the source language.

• Transcribe an audio file:

client("tests/jfk.wav")

• To transcribe from microphone:

client()

• To transcribe from a RTSP stream:

client(rtsp_url="rtsp://admin:admin@192.168.0.1/rtsp")

• To transcribe from a HLS stream:

client(hls_url="http://as-hls-ww-live.akamaized.net/pool_904/live/ww/bbc_1xtra/bbc_1xtra.isml/bbc_1xtra-audio%3d96000.norewind.m3u8")

Enable per-word timing and confidence scores in transcription segments:

client = TranscriptionClient(
  "localhost", 9090,
  word_timestamps=True,
)

When enabled, each segment in the WebSocket response includes a words array:

{
  "segments": [{
    "start": "0.000", "end": "2.500", "text": "Hello world",
    "words": [
      {"word": "Hello", "start": "0.000", "end": "0.800", "probability": 0.95},
      {"word": " world", "start": "0.900", "end": "2.500", "probability": 0.88}
    ]
  }]
}

Boost recognition of specific terms (product names, acronyms, domain jargon):

client = TranscriptionClient(
  "localhost", 9090,
  hotwords="WhisperLive,TensorRT,OpenVINO",
)

The hotwords parameter is a comma-separated string passed directly to faster-whisper's keyword boosting. Also available in the REST API via the hotwords form field.

Real-time speaker identification using pyannote.audio embeddings (optional dependency):

pip install pyannote.audio

client = TranscriptionClient(
  "localhost", 9090,
  enable_diarization=True,
  max_speakers=4,
)

When enabled, completed segments include a speaker field:

{"start": "0.000", "end": "2.500", "text": "Hello", "speaker": "SPEAKER_00", "completed": true}

Diarization uses online cosine-similarity clustering of speaker embeddings. If pyannote.audio is not installed, the server logs a warning and continues without diarization.

The OpenAI-compatible REST endpoint also accepts known_speaker_names and uploaded known_speaker_references multipart fields. When speaker fields are supplied with response_format="verbose_json", segments include a speaker field.

Batch multiple client sessions into single GPU calls for higher throughput:

python3 run_server.py --port 9090 --backend faster_whisper \
  --batch_inference --batch_max_size 8 --batch_window_ms 50

Accept raw PCM int16 audio from clients (useful for embedded devices):

python3 run_server.py --port 9090 --backend faster_whisper --raw_pcm_input

Audio is automatically normalized to float32 range [-1.0, 1.0]. Clients can also set audio_format in the initial websocket options to float32 (default), int16, or uint8.

StreamingTranscriptionClient lets you push raw PCM audio bytes from any source — a live microphone capture loop, a network stream, an audio pipeline — and receive transcripts via callbacks as speech is detected. Unlike TranscriptionClient, it does not manage audio capture internally; you control when and how audio is fed.

A runnable example that reads from an audio file and streams the chunks is at examples/manual_audio_chunking.py:

python examples/manual_audio_chunking.py --file assets/jfk.flac

Example usage:

from whisper_live.client import StreamingTranscriptionClient

client = StreamingTranscriptionClient(
    "localhost", 9090,
    lang="en",
    model="small",
    on_session_started=lambda: print("Server ready"),
    on_partial_transcript=lambda text, segs: print(f"… {text}", end="\r"),
    on_committed_transcript=lambda text, segs: print(f"✓ {text}"),
    on_error=lambda e: print(f"Error: {e}"),
    on_close=lambda: print("Closed"),
)

with client:
    for chunk in my_audio_source:          # any cadence, any chunk size
        client.send(chunk, pcm_format="int16")

Audio must be mono, 16 kHz PCM. Two formats are accepted:

NumPy arrays can be sent with send_array():

import numpy as np
samples = np.frombuffer(raw_bytes, dtype=np.int16)
client.send_array(samples)

Callbacks

• Run the server with your desired backend as shown here.
• Transcribe audio directly from your browser using our Chrome or Firefox extensions. Refer to Audio-Transcription-Chrome and https://github.com/collabora/WhisperLive/blob/main/TensorRT_whisper.md

Aavaaz is a production-grade speech-to-text platform built on WhisperLive, adding speaker diarization, sentiment analysis, PII redaction, webhooks, and OpenAI-compatible REST APIs. It's free to try -- see the Aavaaz README to get started.

Use WhisperLive on iOS with our native iOS client.
Refer to ios-client and ios-client/README.md for setup and usage instructions.

• GPU

  • Faster-Whisper

  docker run -it --gpus all -p 9090:9090 ghcr.io/collabora/whisperlive-gpu:latest

  • TensorRT. Refer to TensorRT_whisper readme for setup and more tensorrt backend configurations.

  docker build . -f docker/Dockerfile.tensorrt -t whisperlive-tensorrt
  docker run -p 9090:9090 --runtime=nvidia --gpus all --entrypoint /bin/bash -it whisperlive-tensorrt
  
  # Build small.en engine
  bash build_whisper_tensorrt.sh /app/TensorRT-LLM-examples small.en        # float16
  bash build_whisper_tensorrt.sh /app/TensorRT-LLM-examples small.en int8   # int8 weight only quantization
  bash build_whisper_tensorrt.sh /app/TensorRT-LLM-examples small.en int4   # int4 weight only quantization
  
  # Run server with small.en (pick one engine)
  python3 run_server.py --port 9090 \
                        --backend tensorrt \
                        --trt_model_path "/app/TensorRT-LLM-examples/whisper/whisper_small_en_float16"
  # or int8 / int4:
  # --trt_model_path "/app/TensorRT-LLM-examples/whisper/whisper_small_en_int8"
  # --trt_model_path "/app/TensorRT-LLM-examples/whisper/whisper_small_en_int4"

  • OpenVINO

  docker run -it --device=/dev/dri -p 9090:9090 ghcr.io/collabora/whisperlive-openvino
  

  • AMD ROCm (faster-whisper on AMD GPU via CTranslate2 ROCm wheel)

  docker build -f docker/Dockerfile.rocm -t whisperlive-rocm .
  docker run --rm -it --device=/dev/kfd --device=/dev/dri \
      --group-add "$(getent group video | cut -d: -f3)" \
      --group-add "$(getent group render | cut -d: -f3)" \
      -p 9090:9090 whisperlive-rocm

• CPU

  • Faster-whisper

  docker run -it -p 9090:9090 ghcr.io/collabora/whisperlive-cpu:latest

On macOS, especially on Intel Macs, faster_whisper/ctranslate2 can conflict with OpenMP runtimes loaded by other Python packages. If the server aborts with a duplicate OpenMP runtime error, run the server with KMP_DUPLICATE_LIB_OK=TRUE:

KMP_DUPLICATE_LIB_OK=TRUE python3 run_server.py --port 9090 \
                      --backend faster_whisper \
                      --max_clients 4 \
                      --max_connection_time 600 \
                      --no_single_model

This workaround is intended for local development and testing. For production deployments, prefer using a clean environment that loads only one OpenMP runtime.

• Add translation to other languages on top of transcription.

• Transforming speech technology with WhisperLive
• WhisperFusion: Ultra-low latency conversations with an AI chatbot powered by WhisperLive
• Breaking language barriers 2.0: Moving closer towards fully reliable, production-ready Hindi ASR which is used in WhisperLive for hindi.

We are available to help you with both Open Source and proprietary AI projects. You can reach us via the Collabora website or vineet.suryan@collabora.com and marcus.edel@collabora.com.

@article{Whisper
  title = {Robust Speech Recognition via Large-Scale Weak Supervision},
  url = {https://arxiv.org/abs/2212.04356},
  author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  publisher = {arXiv},
  year = {2022},
}

@misc{Silero VAD,
  author = {Silero Team},
  title = {Silero VAD: pre-trained enterprise-grade Voice Activity Detector (VAD), Number Detector and Language Classifier},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/snakers4/silero-vad}},
  email = {hello@silero.ai}
}