| --- |
| title: Ringg Parrot STT V1 |
| emoji: 🦜 |
| colorFrom: pink |
| colorTo: blue |
| sdk: static |
| app_file: index.html |
| pinned: false |
| license: other |
| license_name: Proprietary |
| short_description: Proprietary Hindi-English code-mixed ASR system |
| tags: |
| - speech-to-text |
| - asr |
| - bilingual |
| - english |
| - hindi |
| - code-mixed |
| - audio |
| - transcription |
| - ringg |
| - real-time |
| --- |
| |
| # Ringg Parrot STT V1 |
|
|
| High-accuracy Hindi-English code-mixed speech-to-text for business voice applications. |
|
|
| [](https://huggingface.co/spaces/RinggAI/STT) |
| [](#access-and-availability) |
|
|
| ## Overview |
|
|
| Ringg Parrot STT V1 is a proprietary automatic speech recognition system for Hindi, English, and Hindi-English code-mixed speech. It is designed for real-time voice products, business workflows, and production-grade speech-to-text use cases. |
|
|
| This Hugging Face Space is intended for product evaluation and release information. The model weights, training code, and internal implementation are not open sourced. |
|
|
| ## Access and Availability |
|
|
| - Model weights: Not available for download from this repository. |
| - Source code: Internal implementation is not open sourced. |
| - Production and commercial access: Contact RinggAI at `sales@ringg.ai`. |
|
|
| ## Playground Usage |
|
|
| 1. Open the [Ringg STT Playground](https://ringg.ai/dashboard/stt). |
| 2. Upload an audio file or use the available streaming interface. |
| 3. Submit the audio for transcription. |
| 4. Review the generated Hindi, English, or code-mixed transcript. |
|
|
| ## SDK and Integration |
|
|
| - Python SDK: [ringglabs on PyPI](https://pypi.org/project/ringglabs/) |
| - Pipecat: Highly compatible with Pipecat toolkit using built-in VAD events. |
| - For integration details, refer to the SDK documentation. |
|
|
| ## Streaming Latency |
|
|
| Typical streaming latency is 60-80ms. |
|
|
| ## Benchmark Results |
|
|
| WER stands for Word Error Rate. Lower values indicate better transcription accuracy. |
|
|
| ### Original WER |
|
|
| | Dataset | Ringg | ElevenLabs | Deepgram | Sarvam | |
| |---|---:|---:|---:|---:| |
| | indictts | **11.58** | 16.06 | 13.65 | 15.37 | |
| | commonvoice | **14.30** | 16.59 | 20.04 | 18.21 | |
| | fleurs | 15.20 | **11.99** | 17.14 | 16.00 | |
| | kathbath | **11.78** | 13.24 | 15.93 | 17.53 | |
| | kathbath_noisy | **13.09** | 13.14 | 17.44 | 16.19 | |
| | mucs | 14.55 | **11.69** | 21.97 | 16.72 | |
| | Overall WER | 13.79 | **13.00** | 19.23 | 16.72 | |
| |
| ### Normalized WER |
| |
| | Dataset | Ringg | ElevenLabs | Deepgram | Sarvam | |
| |---|---:|---:|---:|---:| |
| | indictts | **3.94** | 8.52 | 6.93 | 7.84 | |
| | commonvoice | **6.37** | 13.02 | 14.88 | 13.06 | |
| | fleurs | 9.73 | **7.67** | 11.35 | 9.54 | |
| | kathbath | **7.15** | 10.15 | 11.38 | 10.41 | |
| | kathbath_noisy | **8.37** | 10.01 | 12.98 | 11.78 | |
| | mucs | **6.28** | 6.75 | 12.07 | 7.58 | |
| | Overall WER | **7.27** | 8.94 | 12.36 | 9.76 | |
|
|
| ## Features |
|
|
| - Hindi-English code-mixed speech recognition. |
| - Real-time streaming transcription. |
| - File-based transcription for common audio formats. |
| - Low-latency inference for voice products. |
| - Business-focused deployment and integration support. |
| - Compatibility with modern voice-agent pipelines. |
|
|
| ## Supported Inputs |
|
|
| - Languages: Hindi, English, and Hindi-English code-mixed speech. |
| - Recommended audio: Clear speech with minimal background noise. |
| - Sample rate: 16kHz or higher recommended. |
| - Formats: WAV, MP3, FLAC, M4A, OGG, and OPUS. |
|
|
| ## Best Practices |
|
|
| - Use clear audio captured close to the speaker. |
| - Reduce background noise, echo, and overlapping speech where possible. |
| - Use 16kHz or higher audio for best results. |
| - Test representative production audio before deployment. |
|
|
| ## Use Cases |
|
|
| - Voice assistants and AI agents. |
| - Contact center transcription. |
| - Meeting and conversation intelligence. |
| - Voice search and commands. |
| - Subtitling and content workflows. |
| - Accessibility and documentation workflows. |
|
|
| ## Limitations |
|
|
| - Accuracy may vary with noisy audio, overlapping speakers, strong accents, dialect variation, or low-quality recordings. |
| - Performance can vary across domains, speaker profiles, and audio capture setups. |
| - Very long files or unsupported encodings may require preprocessing. |
| - The hosted demo is intended for evaluation and may not reflect every production deployment configuration. |
|
|
| ## Privacy and Data Notice |
|
|
| Audio handling may depend on the selected deployment, integration, and commercial terms. Review RinggAI privacy terms and deployment documentation before using the service with sensitive, regulated, or personally identifiable data. |
|
|
| ## Benchmark Dataset |
|
|
| RinggAI has released the [ASR Benchmarking Open-Source Dataset](TODO-add-dataset-URL), which includes benchmark audio/data and transcriptions generated by Ringg, ElevenLabs, Deepgram, and Sarvam. |
|
|
| ## Links |
|
|
| - Playground: [Ringg STT Playground](https://ringg.ai/dashboard/stt) |
| - SDK: [ringglabs on PyPI](https://pypi.org/project/ringglabs/) |
| - Organization: [RinggAI on Hugging Face](https://huggingface.co/RinggAI) |
|
|
| ## Team |
|
|
| Built by the RinggAI Team. |
|
|
