How to Take Perfect Notes with Your Voice Using AI and Notion

First, I should note that we’ll be building and deploying this automation on Pipedream, which is an automation-builder that is similar to Make.com and Zapier. It’s also my favorite of those platforms.

Here’s a look at how our Pipedream automation will work:

1. When a new audio file is uploaded to Dropbox, Google Drive, or OneDrive, the automation is triggered.
2. The audio is downloaded into your Pipedream account’s temporary storage.
3. We get the duration of the audio.
4. The audio is fully transcribed using one of several available speech-to-text models.
5. We send the transcript to one of several available LLMs to get a summary, title, list of action items, etc.
6. The transcript and LLM response are formatted and checked for errors.
7. We send everything to a new page in Notion.

If you want to store and display the duration of your recording, add the following properties to your notes database:

Then add the following formula in the Duration property’s formula editor:

if(floor(prop("Duration (Seconds)") / 3600) < 10, "0", "") + format(floor(prop("Duration (Seconds)") / 3600)) + ":" + if(floor(prop("Duration (Seconds)") % 3600 / 60) < 10, "0", "") + format(floor(prop("Duration (Seconds)") % 3600 / 60)) + ":" + if(floor(prop("Duration (Seconds)") % 3600 % 60) < 10, "0", "") + format(floor(prop("Duration (Seconds)") % 3600 % 60))
Code language: JavaScript (javascript)

This will take the number of seconds in the audio file and show it in hours:minutes:seconds format (e.g. 00:00:00).

With these properties set, we’ll be able to calculate the full cost and duration in our Pipedream automation, and then send the final numbers to these database properties within our new Notion page.

If you’re unable to add or edit properties in your database, make sure the database isn’t locked:

Unlocking Databases – Ultimate Brain Docs

thomasjfrank.com

And if you want to learn more about Notion formulas, check out my comprehensive formula guide:

Formula Reference Introduction – Thomas Frank

thomasjfrank.com

The workflow links above are my Pipedream referral links. If you’re a power user/team and you decide to upgrade to a paid account, I’ll earn a commission at no extra cost to you.

You don’t need a paid account to run this workflow at all. In fact, I spent hours optimizing it so that the Free Pipedream tier can even handle files up to 8 hours long (and really boosted my coding skills in the progress!).

Why Pipedream?

Pipedream is hands-down my favorite automation platform.

I love it, use it every day, and constantly talk about it online. I use it to build all of my automations. Reasons I love it:

• It is the only platform I know of that allows for npm imports. So not only can you write JavaScript code, but you can use any npm package.
• The free tier is incredibly generous. You can do so much more on Pipedream for free than on Make, Zapier, or other traditional no-code platforms. It’s not even in the same league.
• The team is incredibly responsive and helpful.
• As a creator, I can share workflows with my audience, removing 95% of the setup for you.

It’s a product I am extremely proud to be an affiliate for; in fact, I was the one who asked them to build an affiliate program.

I wanted a way to justify making more tutorials, and to be able to share workflows like this one (which took over 200 hours to build and is actively updated) for free.

Again, you can use this workflow on the free plan – and if you choose the speech-to-text and AI providers that offer free plans, you can use the entire workflow for free! Pipedream’s free plan is so much more generous than those of other tools like Zapier and Make that it’s not even in the same ballpark.

But if you do choose to upgrade, thank you. 🙂

When setting up the Dropbox version of the workflow, follow these steps to set up the trigger:

1. Set your Path to the folder in Dropbox you want to use. You’ll likely need to type to search for it; Pipedream doesn’t pull many folders by default.
2. Ensure Include Media Info is set to True.
3. Ensure Include Link is set to True.

After hitting Save and Continue, you’ll follow the same instructions as detailed outside of toggle for the Google Drive version.

You’ll need to upload an audio file to your chosen Dropbox folder to create a valid test even (do not click the Generate Test Event button), then continue on and set up the Download File to TMP action.

When setting up the Dropbox version of the workflow, follow these steps to set up the trigger:

1. Click Folder from Optional Fields and choose a folder in your OneDrive account. You’ll upload audio files to this folder.

After hitting Save and Continue, you’ll follow the same instructions as detailed outside of toggle for the Google Drive version.

You’ll need to upload an audio file to your chosen OneDrive folder to create a valid test even (do not click the Generate Test Event button), then continue on and set up the Download File action.

If you’re using Dropbox, your download step should be called download_file_to_tmp.

Important: Do not rename this step. The following steps in the workflow expect this step to have that exact name. Changing the name will break the workflow.

Note that while Pipedream does provide a Dropbox → Download File to TMP action, this workflow uses a custom version coded by me. This is due to the fact that (as of this writing), Pipedream‘s official Dropbox download step uses very memory-inefficient code that causes workflow runs to fail. My version uses the more memory-efficient methods that Pipedream‘s own Google Drive and OneDrive download actions use.

To set up your download_file_to_tmp step, set up these fields:

1. Dropbox Account: Add your previously configured Dropbox account.
2. File Path: {{steps.trigger.event.path_lower}} (set for you)
3. File Name: {{steps.trigger.event.name}} (set for you)

Hit Test.

If you’re using Microsoft OneDrive, your download step should be called download_file.

Important: Do not rename this step. The following steps in the workflow expect this step to have that exact name. Changing the name will break the workflow.

The OneDrive version of this workflow uses Pipedream‘s official Microsoft OneDrive → Download File action.

To set up your download_file step, set up these fields:

1. Microsoft OneDrive Account: Add your previously configured Microsoft OneDrive account.
2. File ID: {{steps.trigger.event.id}} (set for you)
3. New File Name: {{steps.trigger.event.name}} (set for you)

Hit Test.

Google’s Gemini models can do speech-to-text natively – you don’t need a specific speech-to-text model.

Due to this, Gemini lists their pricing in token costs, and doesn’t provide per-hour costs for audio. Therefore, it’s difficult to calculate Lowest Paid Tier cost for the service. It’s likely very cheap if you select an older model like Gemini 1.5 Pro.

Note that Google explicitly does train their models on free-tier user input.

Here’s how you’d obtain an API key with Groq. Head to the Groq Console and create an account if you haven’t done so already.

On the API Keys page, click Create API Key.

Give your key a name, then click Submit.

When the key is generated, click Copy. You won’t be able to access this key in the future. It’s a secret; treat it like a password. If you’re on a paid tier, this key can be used to make requests to Groq that will result in charges. Always keep API keys secret; keep them safe!

Back at Pipedream, create a Groq Cloud connection by pasting your key into the API Key field. You can also give the connection a name if you like.

Local transcription and AI models have come a long way since I wrote the first version of this guide in 2022! It’s now possible to build some powerful speech-to-text workflows that work directly on a desktop computer.

I’ll note that local transcription on a mobile device is still not super powerful. There are apps that can do it (such as Hello Transcribe), but Whisper models can be tough on a phone’s battery life, so I don’t yet find them ideal for very long transcriptions.

On a desktop computer, though, you can do a lot. On my Mac computers, I use MacWhisper to both:

• Transcribe long files and create captions (for my YouTube videos)
• Quickly write with my voice in any text field using their Dictation feature (I often use this to take notes in Flylighter, our Notion web clipper!)

On Windows computers, I forked a project called whisper-writer and added a ton of capabilities. My fork allows me to “write with my voice” using the dictation tool, and even has AI cleanup and prompting features. I added the ability to use both remote services and local services; I can even use local Ollama models for AI cleanup, making the entire process 100% free, local, and private.

My fork requires some setup to get working, so it is only recommended for technical folks right now. I also can’t provide any support for it whatsoever (truly zero, you’ve been warned!), but I did want to note that it exists.

This property should always have a value of {{steps}}. It simply provides the Transcribe_Summarize step with access to the return values of the previous steps.

Allows you to select the transcription (speech-to-text) service you want to use. Current options include:

• Groq
• Deepgram
• AssemblyAI
• ElevenLabs
• Google Gemini
• OpenAI

I personally use Groq with the distil-whisper-large-v3-en model, and recommend it for most people who are uploading English-language audio files.

Once you select a service, its corresponding App property will be dynamically loaded. I haven’t included those App properties in this reference.

Allows you to select the AI summary service you want to use. Current options include:

• Groq
• Anthropic
• Google Gemini
• OpenAI
• Cerebras
• None

I personally use Groq with the meta-llama/llama-4-scout-17b-16e-instruct model, and recommend it for most people.

Once you select a service, its corresponding App property will be dynamically loaded. I haven’t included those App properties in this reference.

You can set this to None if you only need transcription. However, you’ll need to select a service here if you want:

• AI-generated title
• Summary
• Summary lists (Main Points, Action Items, etc)
• Custom prompt
• Translation
• AI cleanup

Set this to True in order to unlock several advanced options.

This is set to False by default in order to keep the basic setup from being overwhelming to users who don’t need advanced options.

Allows you to select the specific speech-to-text model that will transcribe your audio file.

Options are dynamically set based on the Transcription Service you’ve chosen. Current options include:

• OpenAI: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe
• Deepgram: nova-3, nova-2, nova
• Groq: whisper-large-v3-turbo, distil-whisper-large-v3-en, whisper-large-v3
• Google Gemini: gemini-2.0-flash, gemini-2.0-flash-lite, gemini-1.5-flash
• ElevenLabs: scribe_v1
• AssemblyAI: best, slam-1, nano, universal

Allows you to select the specific large language model (LLM) that will perform all LLM functions, including summary generation, translation, etc.

Options are dynamically set based on the AI Summary Service you’ve chosen. Current options include:

• OpenAI: gpt-4.1-nano, gpt-4.1-mini, gpt-4.1, gpt-4o-mini, gpt-4o
• Anthropic: claude-3-5-haiku-latest, claude-3-5-sonnet-latest, claude-sonnet-4-0, claude-opus-4-0
• Google Gemini: gemini-2.0-flash-lite, gemini-2.0-flash, gemini-1.5-flash
• Groq: llama-3.1-8b-instant, llama-3.3-70b-versatile, meta-llama/llama-4-scout-17b-16e-instruct
• Cerebras: llama-4-scout-17b-16e-instruct, llama3.1-8b, llama-3.3-70b

If you’ve enabled Advanced Options, you can override your choice here using the Custom AI Model property.

Each option you choose here will instruct your AI Model to generate a specific list – except Summary, which will simply generate a paragraph summary.

Current options include:

• Summary
• Main Points
• Action Items
• Follow-up Questions
• Stories
• References
• Arguments
• Related Topics
• Jokes
• Chapters

Each chosen option here will add a new line to the system instructions that are sent along with each chunk of your transcript. If your transcript is split into multiple chunks, the model will generate a summary and list for each chunk – though all chunks will be merged into singular, final lists.

This property allows you to write a custom prompt that will be applied to the entire transcript.

For example, you could write:

Write a blog post draft based on the transcript. If the transcript is in first person, write in first person. Don’t include headings, but do include bullet list items and numbered list items where applicable.

Or you could do something like:

Generate a strong argument against the argument made in the transcript.

This property allows you to go beyond the more limited options in the Summary Options property.

Note that this prompt will be run once, and your entire transcript will be sent to your AI Model along with it. It will not be applied to transcript chunks.

Since you can write anything in this prompt, my workflow isn’t able to specify quite as rigid of a format for the return values as it can for the summary options. It simply instructs the model to return a markdown string.

Therefore, you may wish to add additional instructions about formatting to your prompt, like I did in the blog post prompt example above.

Finally, note that your custom prompt can be found at {{steps.Transcribe_Summarize.$return_value.custom_prompt}} if it exists. The workflow is designed to continue running if the Custom Prompt step fails, so that you still get your transcript and additional summary lists.

Most transcription services limit the size of uploaded audio files. To deal with this, larger audio files are split into chunks (using ffmpeg), and then those chunks are sent concurrently to the transcription service.

This has the additional benefit of speeding up the workflow. Instead of sending one long audio file in a single request, we can send many small audio files, all at the same time. This allows the transcription service to transcribe the chunks in parallel.

By default, this workflow will split files into chunks of 10mb each, even though the smallest file size limit of any service is 24mb. We do this to gain that speed advantage. You can change this setting to anywhere between 4mb and 24mb; however, chunks will always be limited to 10 minutes in length.

When set to False, audio files will always be split into chunks – even if your chosen Transcription Service can handle large uploads (e.g. Deepgram can handle 2gb files, although this workflow itself has a hard cap of 700mb, and will fail on even smaller files unless the Memory setting is increased).

Set this to True to disable audio file chunking.

When set to True, the Transcribe_Summarize step will not automatically clear the audio file out of Pipedream‘s temporary directory after processing it.

This allows the Send_to_Notion step to upload the audio file to Notion, as long as it is within your workspace’s file upload limits.

Defaults to True.

When enabled, this will downsample your audio file to 16kHz mono and convert it to M4A format (32kbps) before transcription.

You probably don’t need this.

If you’ve downloaded your audio file to temporary storage and know the path to it, you can enter it here. You can use this if you’re downloading your file to temp storage using another type of action step (e.g. another cloud storage service).

Path should start with /tmp/ and include the file name. Example: /tmp/my-audio-file.mp3

If a previous step has provided the entire path to the file, you can reference that step’s path here.

If a previous step has provided the file name, you can reference it like so: /tmp/{{file_name_variable}}

If you have a value here, it will override the default behavior of the workflow, which looks for specific download_file action steps for Google Drive, Microsoft OneDrive, and Dropbox.

If you’ve provided a custom Path to File, you can also provide the link to the file from your trigger step here. If you don’t provide a value here, your final Notion page may not contain a link to the audio file.

When enabled, this will enable debug mode, which will cause this step to return the full JSON objects for each transcript and summary response.

This will increase workflow memory usage, so you should only use it when testing workflow steps manually. In Build mode, workflow steps have far more memory available than the default 256mb that deployed workflows have.

Set this to the stage you’d like to stop at. This setting should only be used for debugging purposes.

Enter an array of key terms that the transcription model may need help with.

If you’re using AssemblyAI as your transcription service with the slam-1 model and an English-language audio file, these terms will be included in the key_terms parameter of the transcription request. No other transcription services or models currently support this feature within this workflow.

If you’ve enabled AI Cleanup, these key terms will be included in the system prompt for the LLM that cleans up the transcript.

You can enter a prompt here to help guide the transcription model’s style. This prompt will only be used if you’ve chosen a Whisper model.

It might be useful to set a prompt here if you are uploading files in a non-English language. Occasionally, Whisper models will automatically translate non-English audio into English text. While you can more strongly prevent this by setting a specific language in the Transcription Language property, writing a prompt here in your chosen language may also help.

You might want to use this instead of the Transcription Language property if you’re hoping to use one instance of this workflow for audio files in multiple different languages. This is because if you set a transcription language, the model will always try to output text in that specific language.

Set the temperature for the transcription model. Valid values are integers between 0 and 20 (inclusive), which are divided by 10 to achieve a final value between 0 and 2.0.

This value is only used for Whisper models.

This property sets the language for the transcription model. It instructs the model to output text in the specified language.

If you’re uploading non-English audio and you’re having problems with the model translating your audio into English, you can set a specific language here to prevent that from happening.

If you set a language in this property, your transcript and your chosen summary options will be translated into that language, if it differs from the original language of the transcript.

You’ll only see this property if you have set an AI Service, as a LLM is needed to do the translation.

If you do translate your transcript into another language, the workflow will provide the original language transcript as well.

This option will use your chosen AI model to clean up the transcript. If you’ve provided an array of terms in the key terms property, those terms will be included in the system prompt for the AI model that cleans up the transcript, which you can help with ensuring proper spelling and capitalization for proper terms.

This option sets the maximum number of paragraphs for each chunk of your transcript, and therefore the max number of paragraphs that will be sent to your chosen AI service in each summarization request.

If you set a smaller number here, that’s going to result in a more dense summary, as the same summarization prompt will be run for a smaller chunk of the transcript, and hence more requests will be made as the transcript will be split into more chunks.

Splitting the transcript into chunks will also slightly speed up the workflow as each chunk will be sent to the AI service in parallel except for the first chunk. It is sent ahead of time allowing its summary to be included in the system instructions for all other chunks in order to improve summary quality.

By default, this option will split your transcript into 20 paragraph chunks, where each paragraph is three sentences each.

You can set this property as low as one paragraph per chunk and as high as 5000 paragraphs per chunk, which is effectively the limit of the models themselves, given that current large language models will likely not be able to handle 5000 paragraphs of text.

This option sets the verbosity of your summary and the generated lists for each transcript chunk. It defaults to Medium.

Here’s what each setting does:

• High – Summary will be 20-25% of the transcript length. Most lists will be limited to 5 items.
• Medium – Summary will be 10-15% of the transcript length. Most lists will be limited to 3 items.
• Low – Summary will be 5-10% of the transcript length. Most lists will be limited to 2 items.

Sets the temperature for the AI model. Valid values are integers between 0 and 20 (inclusive), which are divided by 10 to achieve a final value between 0 and 2.0.

If you’d like to use a custom AI model, you can enter the model name here. This will override the default model for the AI service you’ve chosen.

This feature is experimental. You must provide the model name exactly as it appears in the AI service’s API documentation. This feature will only work for models provided by your chosen AI Service.

Example: gemini-2.5-flash-preview-04-17

You’ll still need to select a default AI Model as well.

The Pipedream app property for your Notion account. They’ll use this to connect your Notion account and workspace to Pipedream.

Use this property to select the Notion database in which your new note pages will be created.

If you don’t see the database you’re looking for, refer to the troubleshooting section below.

This property should always have a value of {{steps}}. It simply provides the Send_to_Notion step with access to the return values of the previous steps.

This property lets you choose the sections from the Transcribe_and_Summarize step that you’d like to include in the Notion page.

This property gives you a lot of control over how you use this particular Notion step. For example, you can use this step to only create a new page with the transcript and reference additional sections in new action steps that you add yourself to this workflow.

Select the Title property for your notes. A Notion database will always have one Title property, so you should only ever see a single choice here.

By default, it is called Name.

Choose the value for your note title. Defaults to an AI-generated title based off of the first summarized chunk from your transcription. You can also choose to use the audio file name, or both. If you pick both, the title will be in the format “File Name – AI Title”.

Note: If you didn’t set an AI Service in the Transcribe_Summarize step, your title will be the audio file name even if you choose “AI Generated Title” here. Without an AI Service, the previous step is unable to generate a title.

Advanced: You can also construct a custom title by choosing the Enter a custom expression tab and building an expression that evaluates to a string.

Select the duration property for your database. This must be a Number-type property. Duration will be expressed in seconds.

Note: You can easily create a formula-type property to format the duration into HH:MM:SS format. See this example for the formula code.

If you don’t see an option here, it means your Notion database does not have a Number-type property.

Choose a Select-type property for tagging your note (e.g. tagging it as “AI Transcription”).

Note: This option only supports Select-type properties. If you want to set other property types, such as a Multi-Select or Status property, add a Notion → Update Page step after this step. Set the Page ID value to the {{steps.Send_to_Notion.$return_value.apiResponse.id}} path.

If you don’t see an option here, it means your Notion database does not have a Select-type property.

Choose the value for your note tag. You can choose an existing select value from your database or have this workflow automatically create a new custom value that you enter here.

Choose an emoji to use as the icon for your note page.

Defaults to 🤖. If you don’t see the emoji you want in the list, you can also simply type or paste it in the text box.

Select a date property for your note. This property will be set to the date the audio file was created.

If you don’t see an option here, it means your Notion database does not have a Date-type property.

Select a text-type property for your note’s file name. This property will store the name of the audio file.

If you don’t see an option here, it means your Notion database does not have a Text-type property.

Select a URL-type property for your note’s file link. This property will store a link to the audio file.

If you don’t see an option here, it means your Notion database does not have a URL-type property.

If you’ve used the Custom Section Prompt option in the Transcribe_Summarize step, you can link it here.

Click the field, click Enter a Custom Expression, then find steps → Transcribe_Summarize → custom_prompt.

Alternatively, you can reference a variable from any other custom step you’ve added. The variable’s content must be a string containing plain text or Markdown text.

Set a title/heading for your custom section. Defaults to Custom Section.

If true, this step will attempt to upload the audio file to Notion.

Note: If your workspace is on the free plan, file uploads are limited to 5MB. If your file is larger than this, it will not be uploaded.

Select a Files-type property for your note’s uploaded file. This property will store the uploaded audio file.

You may also leave this blank and solely create an Audio block on the page using the property below.

If you don’t see an option here, it means your Notion database does not have a Files-type property.

If set to True, this step will create an Audio block at the top of the page containing the uploaded audio file.

Select the sections for which you’d like to use toggle heading blocks. These will be toggled closed by default on the page, making the overall page length shorter until you open the toggle that you want.

If True, content in the Transcript and Original-Language Transcript (if present) sections will be compressed into as few blocks as possible, resulting in far fewer calls to the Notion API.

When disabled, each paragraph in the Transcript and Original-Language Transcript sections will be its own block.

This option can be useful if you’re transcribing very long audio files. If you’re running into timeout issues with this Send to Notion step, try setting this to True.

If True, content in the Timestamped Transcript section will be compressed into as few blocks as possible, resulting in far fewer calls to the Notion API.

When disabled, each paragraph in the Timestamped Transcript section will be its own block.

This option can be useful if you’re transcribing very long audio files. If you’re running into timeout issues with this Send to Notion step, try setting this to True.

Set this to True to activate a bunch of additional settings. Defaults to False.

If set, the Compress Transcripts and Compress Timestamps options will only be used if the section’s content contains more than this many blocks.

Example: If set to 100, only a transcript that originally has more than 100 blocks will be compressed.

Leave this blank or set it to 0 to disable the threshold.

Choose the type of header block to use for each section. Defaults to Heading 1.

If set, the order of the sections in the note will be as specified in the object. Keys must be the section names, and must match the section names shown below. Values must be the section order number, starting with 1. Defaults to the canonical order of sections, which is the following:

• summary
• transcript
• original_language_transcript
• vtt
• main_points
• action_items
• follow_up
• stories
• references
• arguments
• related_topics
• chapters
• custom_section

Example: If you wanted your Custom Section at the top of the page, add an entry with custom_section as the key (left text box) and 1 as the value (right text box).

If True, the step will not create a Callout block at the top of the page in the case of a failed audio file upload.

If True, the step will construct the Notion page object, but will not send it to Notion. Instead, it will return the constructed object itself. This can be useful for testing and debugging.

This workflow can handle very long files. I’ve successfully used it to transcribe an 8-hour public-domain audiobook recording of The Illiad in testing, using the default 256mb Workflow Memory setting.

There are two primary limiters to the length of audio files this workflow can handle:

1. Workflow Memory
2. Notion API speed

Pipedream runs on top of AWS Lambda, which means that the “Workflow Memory” setting (which defaults to 256mb) also controls the amount of CPU power the workflow gets.

Through extensive testing, I’ve learned that the ffmpeg process that splits the audio file into chunks it extremely memory-efficient, but it can get hindered by limited CPU resources when handling large files. Things get even worse when a file is large and long in duration – though the reasons aren’t clear to me.

If you need to transcribe extremely long files, you’ll likely need to increase the Workflow Memory setting. You may be able to squeeze extra performance from the default 256mb setting by compressing your audio file (e.g. using a local ffmpeg command like ffmpeg -i file.mp3 -vn -c:a libmp3lame -ar 16000 -b:a 32k file_compressed.mp3 – assuming the input file is called file.mp3), but this can only go so far.

The other limitation is the Notion API. When transcribing audio or running LLM operations, we can split the audio file or transcript into chunks, and concurrently run operations on each chunk.

However, when we’re creating the final Notion page, we have to make API calls sequentially. When you make concurrent API calls (multiple at once), you can’t be sure that they’ll be completed in the correct order. Some may fail and require retries, or may just be completed more slowly.

For this reason, the Send to Notion step must make a request, get confirmation that it succeed, and only then make the next request. Otherwise, you’d run the risk of seeing your transcript out of order.

The limitation here is the Workflow Timeout setting. By default, Pipedream workflows will timeout after 30 seconds if they haven’t finished. You can bump up this limit:

• 5 minutes max on the free plan
• 12.5 minutes max on paid plans

Keep in mind that each increment of 30 seconds will cost 1 credit (for each 256mb of Memory you’re using – e.g. 60 seconds + 1024mb memory = 8 credits).

I’ve set a hard cap of 700mb for file uploads to this workflow. That said, unless you change the default workflow memory setting, the workflow will likely fail, even on smaller files.

In my testing, I found the default 256mb Memory setting will easily handle files up to around 150mb. After that, things start getting a bit dicey.

There also seems to be a relationship between the size of the audio file and the length of the audio file, which I believe has to do with the amount of work FFmpeg has to do to split longer audio files up into chunks, regardless of their size.

This is an area where further testing would be needed for it to be fully understood, and it seems that, so far, nobody on the internet has done this type of testing.

The upshot is that if you would like to upload very large audio files to this workflow, you’ll likely need to increase the workflow Memory setting. Be aware that each increment of 256mb that you add will cost one additional credit per 30 seconds of compute time.

You can learn about Pipedream’s credit model here.

Supported audio file formats depend on the transcription service you choose. In general, this workflow will handle all the typical formats, including mp3, m4a, acc, ogg, wav, flac, etc.

If you upload a file with an unsupported type, the workflow will attempt to convert it to an m4a file before sending it off for transcription. However, the workflow currently will not upload the converted file to Notion.

In general, I recommend uploading files that have a supported file type.

Each time the Transcribe_and_Summarize step runs, it deletes the downloaded audio file from Pipedream’s /tmp/ directory once it’s no longer needed. (It does not delete the file from your cloud storage app.) This is good practice, as it keeps your /tmp/ directory from filling up.

Additionally, Pipedream itself clears out the /tmp/ directory occasionally.

If you’re testing the workflow and hit an Audio file not found error, it means the file has been cleared from the /tmp/ directory one way or the other.

The fix is simple; just re-test your Download File step, hit Continue on that step, then test Transcribe_and_Summarize again.

Additionally, don’t rename the workflow steps. The Transcribe_and_Summarize step expects the Download File steps to use their default names. They should be:

• Google Drive → Download File: download_file
• Microsoft OneDrive → Download File: download_file
• Dropbox → Download File to TMP: download_file_to_tmp

These are the step names that Pipedream sets by default for these actions, and they’re hard-coded into the Transcribe_and_Summarize step’s logic.

If you want to rename steps, you’ll need to enabled Advanced Options in the Transcribe_and_Summarize step, then set the Path to File and File Link properties manually.

If you’ve connected your Notion account, but your desired notes database isn’t showing up in the Notes Database field, here’s the fix.

1. Navigate to your desired Notion database (go to the actual database, not a page that contains it or a linked view of it)
2. Click the ••• menu in the top-right corner.
3. Under Add Connections, find and add Pipedream.

If you’re using Ultimate Brain, this screenshot shows the location of the All Notes database: Ultimate Brain → Archive → All Notes [UB].

Once that’s done, head back to Pipedream and refresh the page. Now your database should show up as an option:

If you’re working with a different database, here’s a trick for making sure you actually navigate to the database itself.

Open any page within that database (e.g. an existing note), then make sure it’s open as a full page (not Side Peek or center modal).

In the breadcrumbs, you’ll see a link to the database directly to the left of the current page’s title. Click that, and you’ll find yourself at the database.

If your workflow runs too long, it may hit a timeout error and fail. By default, Pipedream workflows are set to time out after 30 seconds.

If you’re on the free plan, you can increase this to 300 seconds (5 minutes), which is my recommendation. If you’re on a paid plan, you can increase it as high as 750 seconds (12.5 minutes)

Workflow runs will only consume more credits if they actually take longer, so simply increasing your timeout setting won’t automatically make workflows cost more credits. But it can prevent longer files from failing due to timeouts.

One of the biggest causes of increased workflow time is slow responses from the Notion API. The API doesn’t always respond at the same speed, so it’s good to have a higher timeout setting to compensate.

If you try to upload an audio file that’s too big or too long, you may get an Out of Memory error. I’ve found that I sometimes hit this when uploading files larger than 200mb when my workflow is set to the default 256mb Memory setting.

Pipedream runs on AWS Lambda, which ties workflow memory to CPU power. When you increase the Memory setting, you get more CPU power as well.

This can be helpful for handling very large audio files. The workflow is already extremely memory-efficient, but if you haven’t disabled chunking, ffmpeg needs to split the file into chunks, and bigger files take more CPU power for ffmpeg to handle.

If you get an Out of Memory error, try the following:

• Compress your audio file before you upload it. I had to bump my workflow memory up to 1024mb (1gb) to handle an 8-hour recording of The Iliad at its normal 283mb.
  • But by running the command ffmpeg -i iliad.mp3 -vn -c:a libmp3lame -ar 16000 -b:a 32k iliad_compressed.mp3 in my local computer’s terminal (with ffmpeg installed already), I compressed it down to 106mb.
  • My workflow was able to handle that 106mb compressed version just fine, even though it’s the same length. This is how I discovered that Pipedream workflows need additional CPU power for larger files.
• You can also simply increase your workflow’s Memory setting. Keep in mind that this will increase the number of credits you use per workflow run.
  • I’ve noticed that even some smaller files will cause Out of Memory errors on the 256mb setting, so you may need to bump it up.
  • Alternatively, can try using a transcription service like Deepgram or AssemblyAI. These services don’t have free tiers, but they can handle much larger uploads. This can allow you to disable chunking, which is the main reason workflow runs will hit Out of Memory.

I originally released this tutorial in April 2023 with this video:

The version of the workflow shown in this older video had way lower capabilities, less customization, was more error-prone, and required tons of manual setup.

Therefore, I highly recommend using the new versions I’ve shared above. They take care of nearly everything for you.

If you want to access to old tutorial, however, you can find an archived version here.

You might be wondering if you can build a similar workflow on another automation platform, such as:

• Zapier
• Make.com
• n8n
• Relay

The answer is yes! However, none of these platforms will allow you to make it anywhere near as capable.

Pipedream’s superpower is that it offers developers much more low-level control over the workflows we build. A huge part of that is support for nearly any NPM (Node Package Manager) package, which allows you to bring in powerful libraries and packages that can enhance your own code.

It’s due to NPM support, along with the ability to write complex, multi-file code steps, that I’ve been able to optimize this workflow and add so many features. NPM support also means I can use my own notion-helper JavaScript library, which is the only tool I know of that automatically handles all of the Notion API’s many limitations for you. Without it, handling long transcripts is quite difficult.

I’ve also been able to write code to give you options that would normally require paid “if/then” routing features in most no-code apps. In truth, Notion Voice Notes is closer to a custom app than a no-code workflow.

That being said, you can build a slower, more barebones version on other platforms that may still meet your needs. The overall design of the workflow is simple:

• Trigger when a new file hits a specific folder in your cloud storage app
• Download that file
• Upload it to a speech-to-text service
• Upload the response to an AI service for summarization
• Combine the transcript and summary, then send them to Notion

I use Pipedream for all my automations, so you’ll likely need to seek out another creator or consultant if you need help building this on another service.

If you have an idea for a feature or improvement, open a Github issue about it so we can discuss it!

I’ve done my best to add as many commonly-requested features as possible in the latest build, and I’ve also architected the latest build in a way that makes it easy to add new features.

That said, Notion Voice Notes is a passion project for me, and I can only work on it when I have free time and the desire to do so. I may not always be able to accommodate feature requests in a timely manner, or at all.

However, all of the code that powers my Pipedream actions is source-available, and I hereby give you permission to use and modify it for personal/internal company use.

If you make modifications, you can publish them as your own custom Pipedream actions, which become available for use in workflows within your Pipedream account. To get started:

1. Follow the Action Development quickstart guide
2. Install the Pipedream CLI and publish your actions with it (these actions span multiple files, so the CLI is required to publish them. You can’t use the in-app editor for them)