Voice notes turned into price quotes
Customers send price requests as voice notes, and someone listens to each one and types it into the price calculator. This listens, fills in the calculator and saves the quote.
$600 fixed price Get in touch
No sound. Running on sample data: the companies, people and numbers are made up.
What you get
- A voice note goes in and a priced quote comes out, saved with the transcript and a screenshot of the calculator.
- It uses the price calculator you already have. It opens the page, types the job in and reads the price back, so your prices stay in one place.
- Every quote in one table, priced or waiting for review, with export to a spreadsheet.
- The audio is transcribed on your own machine. Only the text goes to the AI.
What it won’t get wrong
- It won’t guess a missing detail. A note with no collection town goes to a person, and the calculator is never opened.
- It won’t miss a correction. “Three pallets, actually make that five” is priced as five.
- It won’t trust a word it barely heard. A key word the transcription wasn’t sure of stops the quote for review.
- It won’t quote the same note twice. The same audio sent again doesn’t make a second quote.
- It won’t hide a refusal. If your calculator rejects the job, its message is saved in place of a price.
What it needs from you
- Your price calculator, as a web page the bot can open.
- Where the voice notes arrive, and where you want the quotes saved.
- A computer or small server on your side to run it.
- An account with the AI provider, which reads the transcript.
In the video
The sample company is a pallet freight firm. A clean request is priced at £290. A phone-quality note where the caller corrects herself is priced with the corrected numbers. A note with no collection town waits for a person. A job the calculator refuses is saved with the calculator’s message, and a duplicate is skipped.
Want this, or something like it?
As shown here it costs $600, fixed price. That covers setting it up on your own calculator page and your real voice notes, saving the quotes where you need them, such as Google Sheets or Excel, running it on your computer or a small server, and a short guide. If your version is different, you get a fixed price for yours before any work starts. The code is yours.
Thanks, it's on its way.
If I pick your task, I'll email you.
How voice note to quote automation works
Turning speech into text
Speech recognition changed in 2022, when OpenAI released Whisper as open source. It is a model that turns recorded speech into text in many languages, and it is free to run on your own computer. Before that, good transcription mostly meant sending the audio to a cloud service.
The version used here is faster-whisper, a reworked edition that reaches the same accuracy at up to four times the speed and runs on an ordinary processor, with no graphics card. Whisper comes in sizes. The small model is quick. The large one is more accurate on poor audio and about three times slower.
Because the model runs locally, the recording itself never leaves your machine.
What transcription gets wrong
- Numbers and units. “2.4 tons” has to come out as 2.4, in tonnes, and not as 24.
- Names of places and customers the model has never seen.
- Phone-quality audio, wind and traffic.
- People who correct themselves half-way. “Three pallets, actually make that five” is transcribed faithfully, with both the three and the five in it. Something has to work out which one counts.
Whisper also reports how sure it was of each word. That makes it possible to stop when a word that matters, such as the weight or the town, was barely heard.
From a transcript to the fields of a quote
A language model reads the transcript and fills in a fixed form, with the customer, the collection town, the delivery town, the number of pallets, the weight and the service. It resolves the self-corrections, which is the kind of thing language models are good at. It also has to quote the exact words behind each value.
Then ordinary code checks the form against the transcript. Are the quoted words really there? Do they say that number, whether it was spoken as digits or as words? Is the town on the list of places you serve? Does 2.4 tonnes come to 2,400 kg? If anything fails, the note goes to a person with the reason. No price is produced from a guess.
Getting a price out of your own calculator
There are two ways for a program to get a price. If your pricing tool has an API, the program asks it directly. Most small businesses have nothing like that. They have a web page or a spreadsheet where someone types the job in and reads the price off.
So the program does what the person does. It opens the page in a browser of its own, types the values into the same boxes, presses Calculate and reads the result. This is called browser automation. The tool here is Playwright, driving Chrome with no visible window.
Browser automation has a known weakness. If the page changes, a program can type into the wrong box without noticing. Two habits contain that. After typing, it reads every field back from the page and compares it with what it meant to enter. And it keeps a screenshot of the page with each quote, so anyone can see later what the calculator was shown.
The benefit is that your prices stay in one place. The bot never holds a copy of your price list that could go out of date.
Where each box sits on the page is written in a short settings file. A different calculator needs a new settings file and no new program.
Where voice notes come from
WhatsApp is the usual source. Its voice notes are small audio files, and with the WhatsApp API a program can fetch them as they arrive. Others come as email attachments, from a phone system’s voicemail, or dropped into a shared folder. The format rarely matters, since the common ones (m4a, mp3, wav, ogg) are all read.
One note, one quote
The same note often arrives twice, forwarded or sent again. The program fingerprints each audio file and refuses to price one it has already handled.
Each quote is saved in a single step, with its transcript, the checked fields, the price and the screenshot. If the computer stops half-way, the unfinished note is completed on the next start. It isn’t lost and it isn’t priced twice.
When it fits and when it doesn’t
It fits when the requests resemble each other and the details needed are a short fixed list, as with delivery prices, callout fees or standard jobs.
It doesn’t fit open-ended requests that need a conversation. And it doesn’t replace judgement on unusual jobs. Those are exactly the notes it hands to a person.
