新语言语音转文本工具开发咨询：JS/PHP能否实现？求替代方案

Hey there! Let's tackle your two questions head-on—building a speech-to-text tool for a language Google hasn't indexed yet is such a meaningful project, kudos for taking this on!

1. Can you build this tool using only JavaScript and PHP? #

Short answer: Not fully, if you want to handle the core speech recognition logic entirely with these two languages. Here's the breakdown:

• JavaScript shines for frontend voice capture (you can use APIs like the Web Speech API or libraries like Recorder.js to grab audio from users), and you can even run lightweight ML models in the browser with TensorFlow.js. But training a custom speech recognition model for an undocumented language requires processing massive amounts of audio-text paired data—JS isn't optimized for heavy model training, and its ML ecosystem is far less robust than Python's for this kind of work.
• PHP is solid for backend tasks like data storage, user management, or API routing, but it has almost no mature frameworks or tools for building/training speech recognition models. It's simply not designed for intensive machine learning workloads.

That said, you could use JS for frontend audio capture and PHP for backend data handling—but the core speech-to-text model training and inference part would still need to rely on other tools or services. You can't build the full end-to-end tool with only JS and PHP.

2. What services/programs are good alternatives if building from scratch is too hard? #

Since you're already learning Python, you have a perfect foundation to leverage tools built for this exact use case. Here are your best options:

Open-Source Frameworks (Great for Custom Training) #

• OpenAI Whisper: This is a fantastic starting point. Whisper already supports dozens of languages, and you can easily fine-tune it with your own audio-text dataset for your target language. It's Python-based, so it aligns with what you're learning, and it has a simple API for running inference. Even if your language isn't listed out of the box, fine-tuning with your data will give you usable results quickly.
• Kaldi: A powerful, industry-standard open-source speech recognition toolkit. It's more complex than Whisper, but it's highly customizable for building custom acoustic and language models. It has a steep learning curve, but it's ideal if you need full control over the model.
• CMU Sphinx: A lighter, older open-source option that's still useful for small-scale custom language models. It has clear documentation for training new languages, making it a solid choice if you're working with limited resources.

Cloud-Based Custom Speech Services (No Infrastructure Management Needed) #

• AWS Transcribe Custom Language Models: Upload your audio-text paired data, and AWS will handle training a custom speech recognition model for you. It takes care of all the heavy infrastructure and training work, so you can focus on collecting quality data.
• Azure Speech Service Custom Speech: Similar to AWS, Microsoft's service lets you upload datasets to train a model tailored to your language. It also includes tools to test and refine model accuracy over time.
• Google Cloud Speech-to-Text Custom Models: Even though Google hasn't indexed your language natively, their custom model training feature lets you build a recognition model using your own data. It integrates smoothly with other Google Cloud tools if you need backend support.

Hybrid Approach (Leverage Your Existing Skills) #

Combine your JS/PHP expertise with Python for the ML heavy lifting:

• Use JavaScript to build a frontend that captures audio input from users.
• Use PHP to handle user accounts, store audio files, and manage dataset uploads.
• Use Python (with Whisper or Kaldi) to run speech-to-text inference and model training, exposing an API that your PHP backend can call.

This way, you get to use what you already know while leveraging the best tools for the machine learning part.

Pro tip: The most critical piece of this project is collecting high-quality audio-text paired data for your target language. Without enough clean, representative data, even the best tools won't produce accurate results. Start small—collect a few hours of data first, test with Whisper fine-tuning, and iterate from there.

内容的提问来源于stack exchange，提问作者Jamille