Bark 2.0
Speak naturally. Get clean text.
Bark 2.0 is a 2-in-1 speech recognition and cleanup model that turns everyday speech into clear, readable text. It handles filler words and false starts within the speech model itself—no separate cleanup LLM required. We post trained Bark 2.0 from Parakeet Ultra.
Built for fast, private, on-device dictation, Bark 2.0 achieves 3.79% WER on the DisfluencySpeech cleaned-transcription benchmark and processes one minute of audio in just 261 ms on Apple M5 Max with its MLX 4-bit version, measured on short recordings with the model already loaded.
One model. Built-in cleanup. Runs locally.
Built-in cleanup
Speak with natural pauses, fillers and restarts. Bark is designed to produce a cleaned transcript directly, keeping dictation simple without a second text-generation model.
Illustrative examples of the intended cleanup behavior, not quoted benchmark predictions:
| Spoken input | Cleaned output |
|---|---|
| “Um, could you, uh, open the settings?” | “Could you open the settings?” |
| “I want to—I need to change my password.” | “I need to change my password.” |
| “Please send the, uh, send the report tomorrow.” | “Please send the report tomorrow.” |
Speech recognition and cleanup benchmark
Bark 2.0 delivers the lowest cleaned-transcription WER among the systems listed below on this benchmark.
| Model | Runtime / Precision | Hardware | WER ↓ | ms/minute audio ↓ |
|---|---|---|---|---|
| Bark 2.0 | MLX 4-bit | Apple M5 Max | 3.79% | 261 |
| Bark 1.1 | MLX 4-bit | Apple M5 Max | 4.99% | 1,019 |
| Bark 1.1 | ONNX 4-bit | NVIDIA H200 | 4.91% | 1,122 |
| Bark 1.0 | MLX 4-bit | Apple M5 Max | 17.00% | 784 |
| Whisper large-v3 | CUDA BF16 | NVIDIA H200 | 10.02% | 3,573 |
| Whisper large-v3-turbo | CUDA BF16 | NVIDIA H200 | 10.21% | 1,339 |
| Distil-Whisper large-v3.5 | CUDA | NVIDIA H200 | 11.29% | 813 |
| Parakeet TDT v2 | NeMo FP32 | NVIDIA H200 | 15.84% | 486 |
| Parakeet TDT v3 | NeMo FP32 | NVIDIA H200 | 16.57% | 443 |
| Parakeet 1.1B | NeMo FP32 | NVIDIA H200 | 13.92% | 708 |
| Parakeet Redux | Photon BF16 | NVIDIA H200 | 17.97% | 289 |
| Parakeet Ultra | Photon BF16 | NVIDIA H200 | 16.75% | 276 |
| Parakeet Ultra | Community MLX BF16 | Apple M5 Max | 16.71% | 223 |
Dataset and evaluation
Evaluated on the DisfluencySpeech test split: 250 English recordings totaling 27.22 minutes, using identical audio and reference transcripts across systems. The dataset provides both verbatim and cleaned transcripts of disfluent speech.
Bark combines speech recognition with disfluency cleanup. WER here is measured against the dataset's cleaned transcript_c references, which remove fillers and abandoned starts, rather than against verbatim speech.
WER = (word substitutions + deletions + insertions) / reference words × 100. Predictions and references use the same text normalization. Lower is better. Retaining fillers removed from the reference counts as an error, as does removing meaningful content.
This is a task-adapted cleanup model versus off-the-shelf baselines, not an evaluation of unseen-domain generalization. This single-speaker benchmark measures cleaned-transcription performance. The Bark 2.0 row reports the MLX variant; it does not represent an ONNX measurement.
Timing
ms/minute audio measures processing milliseconds per minute of input audio; lower is faster. Measurements use batch size 1 and include audio preprocessing and decoding, excluding initial model loading and warmup. Values are rounded to the nearest millisecond. Compare speeds within the same hardware group; these are short-recording measurements normalized by audio duration, not continuous one-minute recording tests. Bark 2.0's MLX figure is the median of three complete 250-clip passes.
Usage
Use Python 3.11 on Apple Silicon for MLX, or Python 3.12 for the ONNX CPU example. Install into a dedicated Python environment. Download only your chosen variant, the helper and the root license. Replace audio.wav with a local English recording of up to 60 seconds.
Apple Silicon Mac: MLX
Tested with MLX Audio 0.5.3 and MLX 0.32.2. MLX uses the Mac GPU automatically.
python -m pip install mlx-audio==0.5.3 mlx==0.32.2 soundfile==0.14.0 soxr==1.1.0 huggingface-hub==1.32.0
hf download ConwayResearch/bark-0.6B-4bit-v2.0 --include "mlx/*" "infer.py" "LICENSE" --local-dir ./bark
python ./bark/infer.py audio.wav --model ./bark/mlx --engine mlx
ONNX CPU
Tested with ONNX ASR 0.12.0 and ONNX Runtime 1.23.2 on macOS CPU. These commands select the CPU provider. They do not configure CUDA or DirectML; GPU support requires a separately verified runtime/provider setup.
python -m pip install onnx-asr==0.12.0 onnxruntime==1.23.2 soundfile==0.14.0 soxr==1.1.0 huggingface-hub==1.32.0
hf download ConwayResearch/bark-0.6B-4bit-v2.0 --include "onnx/*" "infer.py" "LICENSE" --local-dir ./bark
python ./bark/infer.py audio.wav --model ./bark/onnx --engine onnx
The helper downmixes audio to mono and resamples to 16 kHz, then prints the cleaned transcript. Keep all files in the selected model directory together. Inference runs locally after provisioning. Applications can reuse load_model and transcribe from infer.py to keep the model loaded between recordings; each separate CLI invocation loads it again. The performance table above measures the model-loaded case.
Licensing and attribution are included in LICENSE and in each variant's LICENSE file.
Quantized