Ecsow AI Auto Dialer/Voice Broadcasting Blog

Ecsow’s AI detection does two separate jobs. First it turns speech into text. Then it decides what the text means: a person, a voicemail, a screening assistant, a transfer message, or a phone menu.

The second job is a plain text file. You can edit it with Notepad. That is what this page is about.

Where the file lives

rules.tsv sits in the folder where you installed Ecsow. Open that folder and you will find it next to the program.

There is a second place it can be set: the Rules file field in Tools > Options > Voice Activity Detection > AI Settings. If that field points somewhere else, your edits to the install-folder copy do nothing. Check it once and you will not lose an afternoon to it.

Three columns, five labels

Each line is one rule. Three columns, separated by a single Tab character:

S	4	who is calling
H	4	this is he
VM	5	leave a message after the tone
T	5	please hold while i connect you
IVR	5	press 1

Read the first line as: label S, weight 4, phrase “who is calling”.

Those five letters are the whole vocabulary. Everything else in the file describes how strongly a phrase points at one of them.

Weights, and the number 4

The weight says how much you trust a phrase. Each match adds its weight to that label’s score. The call is only classified once a score reaches 4.

So one weight-5 phrase is enough on its own. Two weight-2 phrases also get there. A single weight-2 phrase does not, and that is on purpose — it stops one loose word from deciding a call.

Keep weights between 1 and 5:

• 5 — a complete sentence that could barely mean anything else, like “leave a message after the tone”
• 4 — a strong phrase, like “this is he”
• 3 — fairly specific
• 2 — common and ambiguous, like “hello”
• 1 — a weak hint you expect to combine with others

If two labels both reach 4, order decides, not size. IVR wins first, then transfer, then screening, then voicemail, then human. A larger score does not override that order.

What the matching actually sees

Rules are compared against cleaned-up text, not what the caller literally said. Before matching, the transcript is lowercased, punctuation becomes a space, and apostrophes are removed rather than expanded.

That last one catches people. “Who’s calling” becomes whos calling, while “who is calling” stays who is calling. Two different strings. If callers say both, add both lines.

Hyphens split words, so “follow-up” arrives as follow up. Chinese characters and Spanish accents are kept as they are, so write them the way they appear in the transcript.

Matching is whole-phrase and looks for complete words. A rule of virtual assistant matches inside “you have reached the virtual assistant for Mark” but not inside “virtual assistants”. Longer phrases are usually the better bet. A single broad word like “assistant” also fires when a human says “can I speak to the assistant?”.

Five steps to change something

1. Back up the file. Copy rules.tsv to rules.tsv.bak. You will want the old one back at least once.

2. Edit with a real editor. Notepad++ or similar, set to UTF-8, with whitespace visible. The Tab characters have to be visible or you cannot tell them from spaces.

3. Add one to three specific phrases. Not thirty. One or three, from calls you actually heard go wrong.

4. Save, then restart Ecsow completely. The file is read once, at startup. Closing the window is not enough — exit the program and open it again. There is no reload button.

5. Test it on a call you can predict. Ring a number you already know the answer on: your own voicemail, a screening line, a bank’s phone menu. Then watch what Ecsow does with it.

A voicemail should get your recording. A screening assistant should get your short reply. A keypad menu should be dropped. If the call lands in the wrong bucket, the rule did not fire — and that is nearly always a Tab problem, an apostrophe difference, or the wrong copy of the file.

Four mistakes that cost the most time

Two more small ones. A # at the start of a line is a comment, but a # in the middle is part of the phrase. And an extra Tab inside a phrase makes the line invalid, so it is skipped without complaint.

You can confirm which file is loaded, too. The Rules file field in AI Settings shows the full path it uses. If that path is not the file you edited, that is your answer.

What the rules cannot fix

A phrase table can only answer “what does this text look like?”. It does not control timing or playback.

If the transcript itself is wrong, no rule will help. Fix the model or the audio first: right language, decent recording, 8 kHz phone audio in, 16 kHz model in.

The rest belong to the application, not the file. A call reported as no human while it is still ringing is a startup-timing issue. A playback timeout is a callback issue. A screening call that waits for a transfer and gives up is policy, not classification.

Tuning without breaking what works

Add a little, test on real calls, then decide. Ten to twenty calls tell you more than one.

Two rules of thumb. Do not raise “hello” or “who is this” to 5 just to fix one case. Both are said by screening assistants too, and you will trade one error for a worse one. And change one rule at a time. Five edits at once tell you nothing about which one worked.

If you have not set the feature up yet, start here. If your calls are not in English, install the matching model first. Rules for Spanish will not help a model that only understands English.

Quick answers

Do I need to edit this file at all?

No. The default rules cover normal English calls. Edit it when you hear a specific phrase go wrong more than once.

Does it matter where I put the file?

Yes. It is read from the install folder, or from whatever path the Rules file field points to in AI Settings. Only one copy is used.

Why did my change do nothing?

Three usual reasons: spaces instead of Tabs, no restart, or the program is reading another copy. Check the Rules file field in AI Settings first.

Can I add rules in Spanish or Chinese?

Yes. Save the file as UTF-8 and keep accents and characters as they appear in the transcript. Chinese needs no spaces between characters unless the transcript has them.

How many rules is too many?

There is no limit, but a long list of single words makes labels fight each other. Complete phrases, fewer lines, higher confidence.

Does a higher weight always win?

No. A score of 4 is the entry ticket, and then the label order decides: IVR, transfer, screening, voicemail, human.

Ecsow’s AI-Enhanced Detection comes with a lightweight English speech model built-in. This is why the model fields may be left empty on a fresh installation.

We also offer free downloadable models for five other languages: Spanish, Arabic, Mandarin Chinese, Japanese, Korean and Vietnamese.

The whole setup only has four simple steps: download the model file, extract it, select the model folder inside AI settings, then restart Ecsow. Most of the waiting time is for the file download, and the process usually takes around 10 minutes.

You only need to install an extra language model if you are handling non‑English calls. If you haven’t turned on AI detection yet, please check our introductory setup guide first.

Step 1: Download the model

Go to the Ecsow download page and find the AI detection models. Six languages are listed.

Each link is a .tar.bz2 file. The “base” models hear more accurately than the “tiny” ones and use more CPU. Pick the language you actually call in — the model only understands one language at a time.

Step 2: Extract the archive to a local folder

ou can use 7‑Zip or Windows’ built‑in extraction tool. .tar.bz2 requires two extraction steps: first unpack the .bz2 file, then extract the .tar file inside it.

Place the extracted files in a simple path, for example: C:\EcsowModels\moonshine-base-es

Tip: Use a short folder path without spaces. This helps avoid loading errors.

Now check the folder. It must contain these three files:

• encoder_model.ort
• decoder_model_merged.ort
• tokens.txt

If any file is missing, you may have opened the outer archive folder. Go deeper into the extracted directory and check again.

Step 3: Point Ecsow to the model folder

Open Tools > Options > Voice Activity Detection and click AI Settings. This is the panel you should see.

Find the setting labelled Moonshine model folder, click Browse, and select the folder you extracted in Step 2.

Ecsow will validate the folder automatically once selected. If you see Moonshine v2 model OK in the status bar at the bottom, the model loads correctly. Save your settings.

You do not need to manually select the language; Ecsow detects it automatically from the model files.

Step 4: Match the rules file to the language

After switching to a new language model, you must use the corresponding rules.tsv file.

The rules file (rules.tsv) is what turns transcribed speech into a decision — person, voicemail, screening assistant, phone menu. The default one matches the English model.

If you point the model at Spanish and leave the English rules in place, the two disagree. The model hears Spanish, the rules expect English, and the classification stops being trustworthy.

So when you change the model, change the rules file to match. Not sure which file fits your language? Ask support before you go live with it. It is a one-line fix once you have the right file.

Step 5: Save settings and restart Ecsow

Click OK, then close and restart Ecsow. The new model loads at startup, so a restart is what makes it take effect.

After the restart, the same status line you saw in step 3 is your confirmation. Then make a short test call and check the detection behaves the way the setup guide describes.

When something looks wrong

None of these need a reinstall. Undo the field, restart, and the default English model comes back.

Quick answers

Do I need any of this for English calls?

No. The English model ships inside the installer. Leave the model field empty and it works.

Where do the model files come from?

The links on the download page point to the public release page of the sherpa-onnx speech project, which builds the Moonshine models Ecsow uses. Nothing is repackaged or resold.

Does the model send my audio anywhere?

No. The speech model runs on the machine running Ecsow. Nothing is uploaded and there is no per-minute charge, whichever model you use.

Which editions can load a custom model?

Both. Ecsow Auto Dialer and Ecsow Voice Broadcast share the same settings panel and the same model folder field.

Should I pick the tiny or the base model?

If both exist for your language, tiny is lighter on CPU and base is more accurate. On an older machine with many lines, start with tiny. If your CPU sits mostly idle while calling, move up to base.

Twelve questions, in the order people ask them.

The short answer: old detection listens to how the call sounds, new detection reads what was said. Everything below is that difference, and why it costs money when you get it wrong.

What is answering machine detection, in one sentence?

Software that decides whether a person or a machine answered, before your agent starts talking.

Every dialer has some version of it. The cheap ones do one trick. The expensive ones stack all of them.

How does the old method decide?

It measures sound, in four ways:

Notice what all four have in common. None of them knows what the person said. They measure timing and tone, then guess.

So why does it get it wrong so often?

Because length and silence do not tell you who answered.

A person who picks up with “Hello, Dave speaking, who is this?” sounds exactly like a long greeting. A voicemail that says “Hi, leave a message” sounds exactly like a short hello. The old method has no way to separate those.

Everyone talks about a slider. What is that?

Because the old method measures one thing, it has one control. Turn it stricter and you hang up on more machines. Turn it looser and you hang up on fewer machines, and on more people.

Here is the part people miss. Both mistakes come from the same measurement. So the slider does not reduce errors. It moves them.

Top right and bottom left are the problem cells. A person with a long, formal greeting looks like a machine. A short voicemail looks like a person.

No setting separates them, because the thing being measured is not the thing that differs. Length is not meaning. Words are meaning.

Which mistake is worse?

Hanging up on a real person, by a wide margin. Roughly thirty times worse.

Take 1,000 answered calls and an error rate of one in ten. That is 100 mistakes each way.

Why is it thirty times worse?

Because one costs a conversation and the other costs an opening.

A lost conversation is about three minutes of talk time. An agent produces roughly thirty minutes of talk per paid hour, so the other thirty minutes go to notes, callbacks and waiting. That means ten paid hours to produce five hours of conversation, which is why 100 lost calls cost about ten agent-hours.

A wasted opening is twelve seconds. One hundred of them is twenty minutes.

Ten agent-hours is roughly $40 in Manila and $180 in Ohio. The wasted openings cost about a dollar.

So most systems are tuned the wrong way round. They protect the agent’s patience because that mistake is loud and annoying, and they pay for the expensive mistake because it is silent.

This is the sentence worth remembering: the call you dropped looks exactly like a no-answer in your report. Same row, same zero. You will never see it.

What does the new method do instead?

It turns the first few seconds into text and reads it. That is the whole trick.

Once you have words, length stops mattering. Once you have words, categories that never existed become possible.

Two right, one right by accident, two wrong.

The two wrong ones are the interesting ones. A screening assistant is a guard standing in front of a live prospect, and the dialer hangs up on the guard. A transfer message is the phone system saying “hold on, I am fetching the person you called”, and the dialer hangs up on that too.

You cannot fix either with a slider, because the old method has no word for “screening assistant”. It only has human and machine.

Is it really just reading words?

Yes. Rules look for phrases — “leave a message”, “you have reached”, “who is calling”, “please hold” — and the first match wins.

There is no mystery in the logic. The hard part is doing it fast and locally, on the same PC that is running the calls.

Does it still get things wrong?

Yes, and here is the list I would want if I were buying.

• It waits about three seconds before your agent speaks. You cannot classify speech you have not heard.
• One-word answers are hard. A bare “Hello?” gives the model almost nothing.
• Phone audio is narrowband. 8 kHz, compressed, often noisy. That limits the input, not the model.
• Accents and rare languages are weaker, especially with the small default model.
• People can sound like voicemail. A receptionist reading a script will do it.
• It uses CPU. On a slow machine, calls queue and classification slows down.

None of that is a reason to keep the slider. All of it is a reason to test before you trust it.

How do I measure my own in twenty minutes?

Do not take a vendor’s accuracy number, mine included. Measure your list.

1. Pick 50 numbers you can identify in advance: 20 that a person will answer, 20 that go to voicemail, 10 with a menu.
2. Dial them with the detector on.
3. Write down four counts: humans connected, voicemails caught, voicemails missed, humans wrongly dropped.
4. Look at the last one first. If humans are dropped, the detector is costing you money however good the other three look.

What is bad about Ecsow’s version?

• Three seconds of delay on every answered call, by default. Faster means less to read, and less to read means more guessing.
• It eats CPU. Beep detection runs on almost nothing. This runs a speech model.
• Six workers is the ceiling in the 32-bit build. Beyond that, calls queue.
• Everything is local. No cloud fallback when the machine is busy, and no bigger model unless you install one.
• You can edit the rules, which means you can also break them.

And the biggest one: better detection does not fix a bad list. If most of your numbers are wrong, you now reach voicemail more efficiently. That is not the same as selling more.

Should I turn it on?

If you run five agents and your list is reasonable, yes. The three seconds are annoying for a day and then nobody notices.

If the machine is old, or you dial thirty lines from one desktop, test it at peak first. Queueing calls is worse than a wrong guess.

And if you are still on a dialer whose only control is a slider, at least point the slider the other way. Protect the live person, not the agent’s patience. The wasted opening costs a dollar. The lost call costs a customer.

How to turn it on, and what each setting does is the companion piece. For why saved time matters at all, the talk-time arithmetic is the place to start.

Anything else?

Can it tell a Google screening assistant from voicemail?

Yes. Screening assistants ask who is calling and why, which is a different sentence from a voicemail greeting. Ecsow can play a short recorded reply and wait for the call to be put through.

Does it work on mobile numbers?

Yes, and mobile is where most of the value is. Mobile greetings are short, which was the hardest case for the old duration-based method.

What is the difference between AMD and AI detection?

AMD measures sound: greeting length, beeps, silences. AI detection transcribes the first seconds and reads the words. AMD has two outcomes. This has five.

How long before my agent can speak?

About three seconds by default. That is the cost of listening first. Shorten the initial listen and the agent speaks sooner, but it has read less, so it guesses more.

It listens to the first three seconds of an answered call, turns the speech into text on your own PC, and decides who picked up.

That is the whole idea. Below is the setup, in three steps, and the five numbers you may want to change.

Open Tools > Options > Voice Activity Detection and tick Enable AI-Enhanced Detection.

That is the entire default setup. A speech model ships inside the installer, so you can leave the model and rules fields empty.

Most dialers sort answers into two buckets: a person or a machine. This sorts five.

The fourth row is the one people notice. “Please hold while I transfer you” used to sound like a machine, so the dialer dropped a warm call. Now it waits.

Why the old method struggled with rows three and four: this explains it in twelve questions.

Dial 50 numbers you can identify in advance: 20 that a person will answer, 20 that go to voicemail, 10 with a menu.

Write down four counts. People connected. Voicemails caught. Voicemails missed. People wrongly dropped.

Read the last one first. If it hangs up on real people, nothing else on the list matters.

Two fixes cover most of it. Dropping real people? Raise Initial silent listen to 4000 ms. Missing voicemails? Raise Voicemail max wait past 20 seconds.

The five numbers

Two checkboxes sit underneath. On timeout leave voicemail decides what happens when it cannot tell: leave a message or hang up. Hangup IVR line automatically ends calls that land in a phone menu. Both are on by default, and both are right for outbound sales.

One more field worth knowing: Screening reply WAV. Google’s call screening asks who is calling and why, and this is the four-to-six-second recording it hears. Keep it honest and keep it short.

What it cannot do

• It waits about three seconds. It cannot read speech it has not heard yet.
• One-word answers are hard. A bare “Hello?” gives it very little to work with.
• Phone audio is narrow. 8 kHz and compressed, which limits every model.
• Accents and rare languages are weaker, especially with the small default model.
• It cannot fix a bad list. Wrong numbers stay wrong.

It also uses CPU, and the checkbox says so. Each worker reads one call at a time, so two workers on five lines means three calls wait their turn. Start at two, raise it by two, and watch the task manager during your busiest hour rather than a quiet morning.

If calls queue at peak, add workers or dial fewer lines. No setting makes a slow machine fast.

The feature is in both editions. If you already run Ecsow it is one checkbox away, and the trial dials 20 contacts per campaign if you want to test it first. Licences start at $10 a month for one line.

Quick answers

Does it send my audio anywhere?

No. The speech model runs on the machine running Ecsow. Nothing is uploaded, and there is no per-minute charge.

How many calls can it read at once?

The worker setting goes from 1 to 6 in the 32-bit build. Each worker handles one call, and extra calls queue briefly. Start at two.

Does it work in other languages?

Yes. Set the language you are calling, such as en, es or fr. Naming it is faster and more accurate than letting the software guess. The auto option works, but it costs extra CPU.

Which editions have it?

Both Ecsow Auto Dialer and Ecsow Voice Broadcast. Same checkbox, same settings panel.