FluidInference
/

parakeet-tdt-0.6b-v3-ov

@@ -1,37 +1,31 @@
 ---
 license: cc-by-4.0
-track_downloads: true
 language:
 - en
 - es
 - fr
 - de
 - bg
 - hr
 - cs
-- da
-- nl
 - et
-- fi
-- el
 - hu
-- it
-- lv
 - lt
 - mt
-- pl
-- pt
-- ro
-- ru
-- sk
 - sl
-- sv
-- uk
 pipeline_tag: automatic-speech-recognition
-library_name: openvino
-datasets:
-- nvidia/Granary
-- nemo/asr-set-3.0
 thumbnail: null
 tags:
 - automatic-speech-recognition
@@ -41,156 +35,86 @@ tags:
 - TDT
 - FastConformer
 - Conformer
 - NeMo
 - OpenVINO
-- Intel NPU
-- hf-asr-leaderboard
-widget:
-- example_title: Librispeech sample 1
-  src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
-- example_title: Librispeech sample 2
-  src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
 base_model:
-- nvidia/parakeet-tdt_1.1b-v3
 ---
-# **<span style="color:#5DAF8D">🧃 parakeet-tdt-1.1b-v3: Multilingual Speech-to-Text OpenVINO</span>**
-## Model Details
-- **Architecture**: Parakeet TDT v3 (Token Duration Transducer, 1.1B parameters)
-- **Input audio**: 16 kHz, mono, Float32 PCM in range [-1, 1]
-- **Languages**: 24 European languages (see below)
-- **Precision**: FP16 (CPU/GPU), INT8 (NPU)
-- **Backend**: OpenVINO 2025.x
-## Performance
-### librispeech Benchmark, English
-```
-================================================================================
-BENCHMARK RESULTS
-================================================================================
-   Dataset: librispeech test-clean
-   Model: parakeet-v3
-   Device: NPU
-   Files processed: 2620
-   Average WER: 3.7%
-   Median WER: 0.0%
-   Average CER: 1.9%
-   Median CER: 0.0%
-   Median RTFx: 23.5x
-   Overall RTFx: 25.7x (19452.5s / 756.4s)
-   Benchmark runtime: 789.8s
-   Normalization: OpenAI Whisper English
-================================================================================
-```
-### FLEURS Benchmark (350 samples per 24 languages)
-<details>
-<summary><b>View all 24 languages</b></summary>
-| Language | WER | CER | RTFx |
-|----------|-----|-----|------|
-| Bulgarian (bg_bg) | 16.76% | 4.66% | 41.7× |
-| Finnish (fi_fi) | 16.81% | 3.68% | 41.5× |
-| Romanian (ro_ro) | 17.51% | 5.89% | 38.9× |
-| Croatian (hr_hr) | 17.76% | 5.84% | 41.0× |
-| Czech (cs_cz) | 18.52% | 5.30% | 43.1× |
-| Swedish (sv_se) | 18.88% | 5.64% | 41.5× |
-| Estonian (et_ee) | 20.78% | 4.90% | 43.4× |
-| Hungarian (hu_hu) | 20.74% | 6.39% | 41.1× |
-| Lithuanian (lt_lt) | 24.55% | 6.66% | 40.4× |
-| Danish (da_dk) | 25.44% | 9.31% | 44.1× |
-| Maltese (mt_mt) | 25.29% | 9.17% | 41.3× |
-| Slovenian (sl_si) | 28.06% | 9.42% | 38.7× |
-| Latvian (lv_lv) | 30.64% | 8.09% | 42.6× |
-| Greek (el_gr) | 42.74% | 14.99% | 37.2× |
-**Average**: 16.98% WER, 5.39% CER, 41.1× RTFx
-</details>
-## Usage
-### Installation
-```bash
-git clone https://github.com/FluidInference/eddy.git
-cd eddy
-# Build with vcpkg (handles dependencies)
-cmake -S . -B build -DCMAKE_TOOLCHAIN_FILE=[vcpkg]/scripts/buildsystems/vcpkg.cmake
-cmake --build build --config Release
-```
-Models auto-download on first run. Cache location:
-- **Windows**: `%LOCALAPPDATA%\eddy\models\parakeet-v3\files\`
-- **Linux**: `~/.cache/eddy/models/parakeet-v3/files/`
-### CLI
-```bash
-# CPU inference
-build/examples/cpp/Release/parakeet_cli.exe audio.wav --model parakeet-v3
-# NPU inference (6-10× faster)
-build/examples/cpp/Release/parakeet_cli.exe audio.wav --model parakeet-v3 --device NPU
-# FLEURS benchmark (all 24 languages)
-build/examples/cpp/Release/benchmark_fleurs.exe "%LOCALAPPDATA%\eddy\datasets\FLEURS" --device NPU
-```
 ## Supported Languages
-🇮🇹 Italian • 🇪🇸 Spanish • 🇬🇧 English • 🇩🇪 German • 🇫🇷 French • 🇳🇱 Dutch • 🇷🇺 Russian • 🇵🇱 Polish • 🇺🇦 Ukrainian • 🇸🇰 Slovak • 🇧🇬 Bulgarian • 🇫🇮 Finnish • 🇷🇴 Romanian • 🇭🇷 Croatian • 🇨🇿 Czech • 🇸🇪 Swedish • 🇪🇪 Estonian • 🇭🇺 Hungarian • 🇱🇹 Lithuanian • 🇩🇰 Danish • 🇲🇹 Maltese • 🇸🇮 Slovenian • 🇱🇻 Latvian • 🇬🇷 Greek
-## Model Architecture
-4-model FastConformer-RNNT pipeline:
-1. **Mel Spectrogram** (preprocessing)
-   - Converts raw audio → 80 mel-frequency bins
-   - 25ms window, 10ms hop length
-2. **Encoder** (FastConformer)
-   - Processes acoustic features
-   - Outputs embeddings every 80ms
-3. **Decoder** (LSTM)
-   - Prediction network with language model
-   - Maintains state across chunks
-4. **Joint Network**
-   - Combines encoder + decoder outputs
-   - Greedy decoding for token prediction
-**Key Features**:
-- LSTM state continuity across audio chunks
-- Token deduplication via 2D search algorithm
-- Batch chunking: 10s windows with 3s overlap
-- Per-token timestamps (80ms granularity) & confidence scores
-**Recommendation**: Use [V2](https://huggingface.co/FluidInference/parakeet-tdt-0.6b-v2-ov) for English-only applications. Use V3 for multilingual support.
-## Limitations
-- **Language Coverage**: Optimized for 24 European languages; performance may degrade for non-European languages or heavy accents.
-- **Noise Robustness**: Best on clean audio; WER increases with background noise.
-- **Streaming Latency**: ~6 seconds with default buffering (configurable).
-## License
-**CC-BY-4.0** - See [LICENSE](LICENSE) for details.
 ## Acknowledgments
-- **Base Model**: NVIDIA NeMo Team for Parakeet TDT architecture
-- **Optimization**: Intel OpenVINO for cross-platform inference
-- **Benchmarks**: Google Research (FLEURS), OpenSLR (LibriSpeech)
-- **Hardware**: Intel Core Ultra NPU acceleration

 ---
 license: cc-by-4.0
 language:
 - en
 - es
+- it
 - fr
 - de
+- nl
+- ru
+- pl
+- uk
+- sk
 - bg
+- fi
+- ro
 - hr
 - cs
+- sv
 - et
 - hu
 - lt
+- da
 - mt
 - sl
+- lv
+- el
 pipeline_tag: automatic-speech-recognition
 thumbnail: null
 tags:
 - automatic-speech-recognition
 - TDT
 - FastConformer
 - Conformer
+- multilingual
 - NeMo
 - OpenVINO
 base_model:
+- nvidia/parakeet-tdt-1.1b
 ---
+# Parakeet TDT 1.1B V3 - OpenVINO
+[![Discord](https://img.shields.io/badge/Discord-Join%20Chat-7289da.svg)](https://discord.gg/WNsvaCtmDe)
+[![GitHub Repo stars](https://img.shields.io/github/stars/FluidInference/eddy?style=flat&logo=github)](https://github.com/FluidInference/eddy)
+OpenVINO-optimized version of NVIDIA's Parakeet TDT 1.1B V3 model for high-performance multilingual automatic speech recognition on Intel NPUs and CPUs.
+## Benchmark Results
+**Hardware**: Intel Core Ultra 7 155H (Meteor Lake) with Intel AI Boost NPU
+**Software**: OpenVINO 2025.x
+### LibriSpeech test-clean (English)
+| Metric | Value |
+|--------|-------|
+| **Average WER** | 3.7% |
+| **Median WER** | 0.0% |
+| **Average CER** | 1.9% |
+| **RTFx (NPU)** | 25.7× |
+| **RTFx (CPU)** | 5-8× |
+| **Files processed** | 2,620 (5.4 hours) |
+### FLEURS Multilingual (24 Languages)
+| Metric | Value |
+|--------|-------|
+| **Average WER** | 17.0% |
+| **Average CER** | 5.4% |
+| **Average RTFx** | 41.1× |
+| **Total samples** | ~15,000+ |
+**Best performing languages** (WER): Italian 4.3%, Spanish 5.4%, English 6.1%, German 7.4%, French 7.7%
+See [BENCHMARK_RESULTS.md](https://github.com/FluidInference/eddy/blob/main/BENCHMARK_RESULTS.md) for complete per-language results.
+## Performance Comparison
+| Implementation | Device | RTFx (Avg) | WER (LibriSpeech) |
+|----------------|--------|------------|-------------------|
+| **eddy (OpenVINO)** | Intel Core Ultra 7 155H NPU | **25.7×** | 3.7% |
+| Parakeet (PyTorch) | Intel Arc 140V GPU | ~20×* | ~2.5%* |
+| **eddy (OpenVINO)** | Intel Core Ultra 7 155H CPU | **5-8×** | 3.7% |
+> **Note**: Benchmarked on HP EliteBook Ultra G1i. eddy NPU is ~1.3× faster than PyTorch on Intel Arc GPU, with lower power consumption. *V3 estimated from V2 benchmark.
 ## Supported Languages
+**24 European languages**: English, Spanish, Italian, French, German, Dutch, Russian, Polish, Ukrainian, Slovak, Bulgarian, Finnish, Romanian, Croatian, Czech, Swedish, Estonian, Hungarian, Lithuanian, Danish, Maltese, Slovenian, Latvian, Greek
+## Usage
+Python usage via ctypes available - see [eddy repository](https://github.com/FluidInference/eddy) for details.
+## Model Details
+- **Parameters**: 1.1B
+- **Architecture**: FastConformer-RNNT (4-model pipeline)
+- **Languages**: 24 European languages
+- **Blank token ID**: 8192
+- **Context window**: 10s chunks with 3s overlap
+- **Features**: LSTM state continuity, token deduplication, per-token timestamps
+## License
+CC-BY-4.0 - See [LICENSE](LICENSE) for details.
+## Links
+- **GitHub**: [FluidInference/eddy](https://github.com/FluidInference/eddy)
+- **Base Model**: [nvidia/parakeet-tdt-1.1b](https://huggingface.co/nvidia/parakeet-tdt-1.1b)
+- **Documentation**: [Benchmark Results](https://github.com/FluidInference/eddy/blob/main/BENCHMARK_RESULTS.md)
 ## Acknowledgments
+Based on NVIDIA's Parakeet TDT model. OpenVINO conversion and optimization by the FluidInference team.