Hello. I'm Rookie from the AmiVoiceAPI development team.
This may be a bit sudden, but do you use any no-code/low-code platforms in your work?

What is a no-code/low-code platform?

A no-code/low-code platform is a development tool that allows you to create apps and systems without any programming knowledge.
By simply using drag-and-drop and templates, you can quickly create business applications and web services. This is extremely useful for resolving the engineer shortage and improving work efficiency.
In this article, we will explain how to efficiently recognize long audio files using the Asynchronous v2 of the AmiVoice API.
When using the AmiVoice API Asynchronous v2 to recognize large audio files, from the perspective of low-code platform developers and users, using a low-code platform offers the following advantages:

For developers,

• By simply combining components on the low-code platform, you can easily integrate AmiVoice's speech recognition capabilities into your workflow.
  
• AmiVoice API is easy to use, and the speech recognition function of AmiVoice API can be used easily without complex implementation, significantly reducing the hurdles to implementation and operation.
  
• Configuration is simple as the necessary parameters are made public. Once the developer publishes the parameters required for recognition, users can start using the service by simply entering the minimum parameters required for authentication.

For users,

• Simply call the workflow URL of the low-code platform and it can be used on both PC and smartphone.

• Preparation is easy, just save the audio file to cloud storage.

For example, if you are having trouble processing large amounts of audio data, how about this solution?

Designing workflows with the low-code platform Dify

The low-code platform Dify is a platform that allows users to create applications even without specialized programming knowledge.
It is a development support platform that enables you to quickly build advanced AI applications and business tools.

It has an intuitive UI that allows you to visually configure the following settings:

• API integration
• Workflow automation
• Data management

This can significantly improve development efficiency.

How to connect Dify with AmiVoice API

The advantage of Dify is that you can freely combine functional parts called blocks to create a "workflow" that suits your business and its purpose. Below is a diagram showing some of Dify's most common blocks and their purpose categories.

In Dify, these blocks are connected with lines on the workflow screen, and data is passed along as processing progresses.

The main blocks used in this article are shown below. The speech recognition workflow is realized by connecting these seven "blocks" with "lines." Each plugin can be obtained from the Dify Marketplace.

• HTTP request: Speech recognition processing with AmiVoice API
• JSON: JSON Process Plugin
• Social media integration: Slack plugin
• Code execution: Python code
• File operations: File Tools plugin
• Logic: IF/ELSE
• LLM: Text to HTML

Advance preparation

These are items that must be prepared outside of Dify. If you use email instead of Slack, you can change the items you need to prepare.

1. AmiVoice API authentication application key

→ Used for HTTP requests in developer flows.

2. Result acquisition notification destination: Slack WEBHOOK_URL

→ Incorporate it into the external service integration of the developer flow.

With all of this in place, users can perform speech recognition tasks with a single click in the Verify flow.

In Dify's workflow, you upload the audio file to be recognized locally, and then orchestrate it on Dify's low-code platform to easily obtain the session ID result. In this case, Slack is used as a notification method to inform the user of the result.

Of course, by taking advantage of the characteristics of low-code platforms, you can also store files on Google Drive or OneDrive and add work units for each Drive to your workflow to share (exchange) data.

What will be notified to speech recognition users via Slack?

Once speech recognition begins, users will receive a message in Slack similar to the one below.

Users who receive this notification can access the URL in the message and enter the obtained session ID to obtain the text of the speech recognition results. If the processing is not yet complete due to reasons such as the audio file being large, a "Processing" response will be returned, so users should wait a while and access the URL again.

Dify Implementation - Creating a Workflow -

Now it's time to become a developer and create a workflow for Dify. Below is the workflow to be created.
All you have to do is follow the wizard and enter the information you prepared in advance, but I'll explain it a little bit.

Workflow 1 - Start speech recognition

Send the target file to AmiVoice to start speech recognition.

START (File selection）：
This block inputs the audio file to be recognized. The type of speech recognition to be used is also set here.

First, the developer obtains the app_key for speech recognition from the AmiVoice Cloud Platform and prepares the slack_api for notifications in advance.
On the other hand, what the user needs to do is very simple; they just need to prepare the audio file they want to recognize and the type of AmiVoice API engine they want to use.

HTTP REQUEST (HTTP request）:
Use the HTTP node to send a speech recognition request to the AmiVoice API. Refer to the AmiVoice API user manual and enter the required parameters specified in the CURL command directly into the HTTP card. For example, in the following CURL command:

curl https://acp-api-async.amivoice.com/v2/recognitions \
     -F u={APP_KEY} \
     -F d="grammarFileNames=-a-general" \
     -F a=@test.wav

Enter the three key parameters u, d, and a in the CURL command into the body of the HTTP card.
Select POST as the method and specify the variables inside Define.

GET FILENAME (Code Execution):
The name of the uploaded audio file can be obtained using a simple function:

JSON Decoding (JSON Process)：
The response from the AmiVoice API is a string containing the session ID in JSON format, so we will place it here to decode it.

SLACK:
Use the slack_url you set up beforehand as the destination. Send the session ID and the name of the audio file used for speech recognition as a message.

The overall flow diagram is as follows:

Workflow 2 - Obtaining text from speech recognition results

The recognition result text is obtained based on the session ID.

START (user input):
Developers must set an app_key in advance to be used for authentication of the AmiVoice API.
All the user needs to do is obtain the recognition result based on the session ID sent to Slack.

HTTP REQUEST (HTTP request）:
You can get the recognition results by sending a request to the specified endpoint using the HTTP GET method along with the session ID.

IF / ELSE (logic):
Depending on the returned status code, the workflow will let you know whether it is processing or not, and once processing is complete, a downloadable txt file of the recognition results will be generated.

SAVE AS FILE (File Tools）:
The acquired text is saved as a file. It contains not only the speech recognition string but also meta information such as the time of speech.

Furthermore, the resulting text can be fed directly into the "useful cards" in Dify's low-code environment, allowing it to be flexibly incorporated into a variety of applications, such as automatic summarization, entity extraction, and monitoring report generation.

The overall flow diagram is as follows:

Application -Attaching LLM to the flow-

Dify comes with a number of "convenient cards" that can be used for business by default.
I'll try using the hottest card right now, "LLM".

By directly passing the recognized text to the LLM node, it is possible to automate summary generation and content analysis, which is expected to significantly improve the efficiency of organizing meeting minutes and generating business reports.

Below are the audio files used and the recognition results.

Audio used:
　Audio data (limited distribution until the minutes are released) | Ministry of the Environment

By passing this speech recognition result along with the prompt to LLM, result analysis can be automatically generated from the LLM node in Dify.

LLM Prompt："Please analyze and organize the speech recognition results and output them in HTML format. Do not output any unnecessary content."

This is the execution result. If you use generative AI on a daily basis, you probably have a general idea of ​​what it looks like.

Conclusion

What did you think of creating a speech recognition workflow using a no-code tool?
This was my first attempt at speech recognition with Dify, and I was able to create something that was practical.

• I have an idea, but development is a pain
• I want to create a convenient tool to save time, but it's difficult to set up a platform.

I realized that this is truly recommended for those who think the same way.

So, if you have any questions about the workflow, please contact us and we'll work together.
We are also always looking for ideas for no-code apps.

AmiVoiceAPI has developed and released a new "Asynchronous API Version 2" that is suitable for large audio files.
Depending on the file, you can expect much faster processing, so please try it out. The manual is here.

Asynchronous API Version 2

In addition to the free tier of 60 minutes of audio per month, AmiVoiceAPI is currently running a limited-time campaign where you can use AmiVoice API for 1,000 hours free of charge.
Please take a look at our website when you have time.

Person who wrote this article

Rookie

I studied computer science at university, particularly image processing. I am currently working at a speech recognition company, where I am trying to catch up on cloud infrastructure and AI application technology.