# Hugging Face Audio MCP > Hugging Face Audio connects audio processing to your AI client via MCP. It provides four tools to handle the full audio lifecycle: transcribe speech from URLs, classify sounds in files, enhance noisy audio quality, and generate speech from text. Use it to analyze, clean, or synthesize any audio stream directly within your agent workflow. ## Overview - **Category:** ai-frontier - **Price:** Free ## Description Hugging Face Audio connects audio processing to your agent via MCP. It gives your AI client four tools that handle the full audio lifecycle: transcribing speech from URLs, classifying sounds in files, cleaning up noisy audio, and generating speech from text. You can analyze, clean, or synthesize any audio stream right within your agent workflow. **`transcribe_audio`** takes an audio file URL and converts any spoken language into plain text, no matter what language it is. **`classify_audio`** analyzes an audio file URL and gives you a structured list of specific sounds it detects, like a dog barking, a car horn, or human speech. **`enhance_audio`** takes a noisy audio file URL and processes it to return a clean version, reducing background interference and making it clearer. **`text_to_speech`** lets you provide text, and it spits out the corresponding synthetic speech audio as a Base64 encoded string, ready to play back. ## Tools ### classify_audio Analyzes an audio file URL and returns a list of specific sounds detected within the file. ### enhance_audio Takes an audio file URL and cleans the audio by removing background noise, improving the overall clarity. ### text_to_speech Generates synthetic speech audio from a given text and returns it as a Base64 encoded string. ### transcribe_audio Converts spoken language from an audio file URL into a readable, plain text transcript. Supports multiple languages. ## Capabilities ### Transcribe spoken language You pass an audio file URL, and the tool returns the full transcript as plain text, regardless of the language spoken. ### Identify sound types You feed the tool an audio file URL, and it outputs a structured list detailing what sounds were detected (e.g., a dog bark, a car horn, or human speech). ### Remove background noise The tool takes a noisy audio file URL and processes it to return a cleaned version, reducing background interference and improving clarity. ### Generate speech from text You provide text, and the tool outputs the corresponding synthetic speech audio encoded in Base64, ready for playback. ### Process multi-stage audio pipelines You can chain tools—for instance, running `enhance_audio` first, then `transcribe_audio`—to perform complex, multi-step analysis on a single file. ## Use Cases ### Analyzing field recordings A wildlife biologist records ambient sounds at a remote site. Instead of spending hours listening for specific animal calls, the agent calls `classify_audio` on the recording. The agent immediately returns a list, confirming the presence of a specific bird species and noting other background noises. ### Improving noisy call center data The QA team gets audio recordings from a noisy call center. They pass the files to `enhance_audio` first, then run `transcribe_audio`. This process delivers a clean, accurate transcript that bypasses the need for manual human review of garbled audio. ### Creating interactive voice assistants You're building a voice bot. When the user asks a question, the agent uses `text_to_speech` to speak the answer. The system then waits for the user's next input, creating a smooth, natural conversational flow. ### Debugging audio system failures A system is failing because it can't understand the input. The agent first runs `classify_audio` to check if the input contains speech at all. If it detects only music, the agent can report the failure reason before attempting a useless `transcribe_audio` call. ## Benefits - **Transcribe speech** using `transcribe_audio`. Instead of manual listening, your agent instantly converts any audio recording into text, letting it process the words immediately. - **Identify sound patterns** with `classify_audio`. You don't just know there's audio; you know *what* is in it. This is critical for building systems that detect specific events, like alarms or voices. - **Clean up noisy inputs** using `enhance_audio`. When raw audio is messy—think wind noise or background chatter—`enhance_audio` gives you a cleaner file, guaranteeing better results from the subsequent `transcribe_audio` call. - **Create synthetic voices** with `text_to_speech`. You can make your agent speak back a response instantly. It takes simple text input and generates the required Base64 audio output. - **Build complex workflows** by chaining tools. You can run `enhance_audio` -> `transcribe_audio` -> `classify_audio` to build a single, multi-stage analysis pipeline on one file. - **Handle multilingual data** via `transcribe_audio`. The tool supports multiple languages, so your agent doesn't break when it hears speech from different regions or people. ## How It Works The bottom line is you call the tool, and the server returns the processed audio or text data directly to your AI client for the next step. 1. Start by calling the tool with the audio file URL and any necessary parameters (e.g., the text for `text_to_speech`). 2. The MCP Server sends the audio data to the Hugging Face backend for processing. 3. Your AI client receives the result: either plain text (from transcription), a classified list, or a Base64 encoded audio blob (from synthesis). ## Frequently Asked Questions **How do I use the `transcribe_audio` tool with Hugging Face Audio?** You pass the URL of the audio file to `transcribe_audio`. The tool returns the full transcript as plain text, and it supports multiple languages, so you don't need a separate language detector. **Is `classify_audio` better than just checking the audio file metadata?** Yes. Metadata only gives file stats. `classify_audio` analyzes the actual content, telling you *what* sounds are present—whether it's a car, music, or a voice—not just the file's properties. **What is the best workflow for noisy audio?** The best workflow is to run `enhance_audio` first. This cleans the noise. Then, you pass the enhanced output to `transcribe_audio` for the highest possible transcription accuracy. **Does `text_to_speech` support different voices or accents?** The tool generates speech audio from text and returns it as a Base64 string. Check the tool's documentation for specific voice parameters, as the core function is text-to-audio generation. **What format does the `enhance_audio` tool use for noisy audio files?** It accepts audio files via a URL. You simply provide the link, and the tool returns the cleaned, enhanced audio data for you to use. **Can I run `classify_audio` on a large number of audio files at once?** Yes, you can process multiple files by calling `classify_audio` repeatedly with different URLs. The tool handles one file at a time, but your agent can loop through a list of URLs. **Does `text_to_speech` require a specific input format for the text?** No, you just give it plain text. The tool generates the speech audio and returns it to your agent as Base64 encoded data. **How does `transcribe_audio` handle different languages and dialects?** It supports multiple languages. You need to specify the language code for accurate transcription, and the tool will convert the speech into text for you.