Arabic Pronunciation Analysis
Budget: $30 – $250 USD
I’m building an ASR-driven application that listens to users through the microphone and instantly evaluates how closely their spoken Arabic matches a native reference. At the core, I need Dynamic Time Warping for temporal alignment and a Goodness of Pronunciation (GOP) module to score each phoneme.
The interface itself should work in both Arabic and English so learners can switch languages for menus, feedback, and error messages. I already have built most of the logic and mostly just need help making sense of some things. I don't need making an app, just help me with the actual functionality so that it works how I intend it to. Essentially, the users of this project would be children, usually with an Arabic base but lacking in speech or conversation skills.
I will have mainly two modules: one where the user reads what's on screen and gets feedback on it, and the other one is free speech, allowing them to have short conversations with an AI "partner". I'm not sure how to do the free speech part yet, but I will be using a genAI API. Any help with this is also appreciated.
The flow is as follows (for first module):
- Speech in through mic
- TTS is generated through Eleven Labs API
- User audio is compared against TTS with the aforementioned algorithms metrics
- Feedback and score
What matters most to me:
• Accurate, low-latency capture from the live mic
• Reliable DTW alignment against stored reference utterances
• GOP-based scoring that pinpoints problem phonemes and outputs an overall pronunciation score
• Clear, bilingual feedback (Arabic/English) showing heat-maps, numeric scores, or text suggestions
• Well-commented source code and a short README that explains how to retrain or expand the reference corpus
If you have prior experience with speech toolkits such as Kaldi, Julius, CMU Sphinx, or a neural backend like Whisper coupled with custom DTW/GOP routines, please showcase it in your approach. A small demo clip proving real-time scoring will be the first acceptance checkpoint, followed by documentation.
NOTE: I'm past the DTW alignment stage, what I need is a more accurate GOP algorithm, please set your bids accordingly.
The interface itself should work in both Arabic and English so learners can switch languages for menus, feedback, and error messages. I already have built most of the logic and mostly just need help making sense of some things. I don't need making an app, just help me with the actual functionality so that it works how I intend it to. Essentially, the users of this project would be children, usually with an Arabic base but lacking in speech or conversation skills.
I will have mainly two modules: one where the user reads what's on screen and gets feedback on it, and the other one is free speech, allowing them to have short conversations with an AI "partner". I'm not sure how to do the free speech part yet, but I will be using a genAI API. Any help with this is also appreciated.
The flow is as follows (for first module):
- Speech in through mic
- TTS is generated through Eleven Labs API
- User audio is compared against TTS with the aforementioned algorithms metrics
- Feedback and score
What matters most to me:
• Accurate, low-latency capture from the live mic
• Reliable DTW alignment against stored reference utterances
• GOP-based scoring that pinpoints problem phonemes and outputs an overall pronunciation score
• Clear, bilingual feedback (Arabic/English) showing heat-maps, numeric scores, or text suggestions
• Well-commented source code and a short README that explains how to retrain or expand the reference corpus
If you have prior experience with speech toolkits such as Kaldi, Julius, CMU Sphinx, or a neural backend like Whisper coupled with custom DTW/GOP routines, please showcase it in your approach. A small demo clip proving real-time scoring will be the first acceptance checkpoint, followed by documentation.
NOTE: I'm past the DTW alignment stage, what I need is a more accurate GOP algorithm, please set your bids accordingly.