How to choose a speech analytics service
Choose a speech analytics service based on a pilot on your own recordings. Check how it transcribes noisy calls, whether it scores against your checklist with a quote from the conversation, whether it sees chats next to calls, whether it writes the result to amoCRM, and where it stores the data. Compare prices against your own monthly volume.
Transcription, keyword search and semantic analysis
Three different products are sold under the name “speech analytics”, and in a demo they look the same: a screen with the conversation text and colored tags. The difference shows when you ask the system a question. Transcription answers what was said. Keyword search, whether a word came up. Semantic analysis, whether they actually agreed on something.
Transcription turns sound into text. It's handy to read instead of listening, and it makes an old conversation easy to find. It draws no conclusions: the head of sales still reviews every call himself, just with his eyes instead of his ears.
Keyword search tags by dictionary. The manager said “discount”: tag. The client said “expensive”: tag. The trouble is real speech. The client says “well, the neighbours got it cheaper”, and a dictionary with the word “expensive” stays silent. A manager who said “delivery” five times may still never have named a date.
Semantic analysis reads the whole conversation and answers the checklist questions: did the manager get the dimensions, handle the objection, set a next step. That's what a sales team needs. Transcription and the keyword dictionary are layers beneath it, and the buying decision should rest on the quality of the analysis.
Services also differ in where the analytics lives: the three types side by side.
| Criterion | Built into telephony | Standalone service | AI review against your criteria |
|---|---|---|---|
| What it gives you | Recordings and, most often, transcripts with keyword search | Transcripts, dictionary tags and ready-made scoring templates | Semantic analysis of every conversation, with a quote next to each score |
| Channels | Only calls made through that telephony | Calls; chats only if the service connects messengers | Calls and chats as one history per customer |
| Checklist | The set of tags is defined by the provider | Your own points often need a synonym dictionary or custom development | Points are written as plain sentences and adjusted to your niche |
| CRM link | The recording lands in the card if the telephony is connected to the CRM | Depends on ready-made integrations: check for your CRM and PBX | The score, a summary and a task appear in the deal card |
| Data | Stored wherever the telephony keeps its recordings | Ask about the data centre and whether transcripts go to a third-party model | The same questions: country of storage, third-party model, data processing agreement |
| Who it suits | You need to find and replay individual calls | Lots of calls, and a ready-made set of tags and reports is enough | You need to see whether the manager reached an agreement, point by point in your process |
Hypothetical example from a gate sale. The dictionary finds “delivery” four times but misses an objection without the word “expensive”. Each finding sits next to the quote and timestamp, so it can be checked in a minute.
Test transcription on your own recordings
In a demo, every speech analytics service plays a clean call: a good microphone, two calm voices. Your calls sound different. The manager talks through a headset in an open-plan office, the client calls from the car or from a site where someone is using a hammer drill nearby.
An example from selling gates. Tuesday, 11:20, a client shouts into the phone: “Opening two-forty by three, sliding, can you make it by Friday?” If the transcription turned “two-forty” into “two hundred and forty”, the whole analysis that follows is built on an error. The analysis can never be more accurate than the text it works from.
Take ten to twenty of your own recordings of varying quality: a quiet client, a noisy background, fast speech, the names of your products and brands. Run them through the service and read the transcripts yourself. Focus on what matters for the deal: dimensions, prices, dates, names, models. Mistakes in filler words can be forgiven.
Check separately how the system splits speech between speakers. If it confuses the manager with the client, it will attribute the client's objection to the manager, and the checklist score will be off. Ask your telephony provider whether calls are recorded in two tracks: with separate channels for the manager and the client, this problem is easier to solve.
The checklist has to be yours
The ready-made checklists in these services are written for an average salesperson: greeted, introduced themselves, identified the need, handled the objection. Your sales have their own key points. In selling gates the main question is the size of the opening, in dentistry it's a convenient appointment time, in wholesale it's batch volume and payment terms.
Ask the service who changes the checklist and how. It's convenient when an item is phrased as an ordinary sentence: “The manager asked whether the client had already had a survey.” It's worse when every item requires building a synonym dictionary or waiting for a fix from the vendor's developer.
Check how the system explains its score. Next to a flag like “no next step set” there should be the quote from the conversation that led to the conclusion. Without a quote, the manager will dispute the score, and the head of sales will go and listen to the recording himself — right back where he started.
We described in detail which items are worth checking in a conversation in our article “Speech analytics: how to review managers' calls”. Start with three to five items you're actually prepared to act on. Nobody reads a thirty-item checklist to the end during a pilot.
Chats alongside calls
A client rarely goes the whole way by phone. First they write on WhatsApp or Avito, then call, then send photos of the site on Telegram. A service that only sees calls rates a fragment of the story and misses that the client asked the key question in a chat the day before.
A scene. Thursday, 21:15, a client writes on Telegram: “Do you offer instalments?” In the morning the manager calls back, talks at length about models and prices, and says nothing about instalments. Judged by the call, everything is fine. The chat shows that the client's question was never answered.
Late enquiries are more common than managers think. In a live GetGate project, 27% of leads come in the evening and at weekends. If the service only analyses calls during working hours, a sizeable share of client conversations escapes any control.
Check which channels the service connects on its own and how chats get into the review: as one conversation together with that client's calls, or as a separate feed. For evaluating a manager, the first option works better. The whole deal history is in one place, and the score takes into account what the client has already been told in the chat.
Telephony, amoCRM and a report for the head of sales
Recordings should reach the service on their own. If an administrator uploads files by hand once a week, within a month the upload turns into “we'll catch up next week”. Ask which telephony systems have a ready-made connection and what to do if you have your own PBX or managers call from their mobiles.
The second link is with the CRM. The review result is useful in the deal card in amoCRM or Bitrix24: a score, a short summary of the conversation, an unanswered question the system spotted. The head of sales opens the deal and sees what happened, without a separate dashboard. It’s good when the service can set a task: “call back, the client is waiting for a quote by Friday”.
We cover setting up the CRM itself and connecting telephony on the page about amoCRM implementation. If your CRM is kept any old how, start there: speech analytics without a link to the deal turns into an archive of conversations that nobody owns.
Look at the report for the head of sales before you buy. Monday, 9:30, the head of sales has fifteen minutes before the team meeting. They need a short list: which deals need stepping in today, which manager is slipping on which point, what clients asked about more often than usual. Nobody opens a forty-column spreadsheet after the first week.
Where recordings are stored: data and Federal Law 152-FZ
A call recording contains the client's voice, name, phone number and the address of the site. All of that is personal data, and you are responsible for it as the data controller, even if the service does the processing. That's why the question “where are the recordings stored?” is asked before the pilot.
When choosing a service, two requirements of Federal Law No. 152-FZ matter; they are summarized in the TAdviser overview of the personal data law (updated in February 2025). Personal data of Russian citizens must be recorded, accumulated and stored using databases located in Russia. The consent to processing must state who the processing is entrusted to.
Ask the vendor direct questions. In which data center and which country are the recordings and transcripts stored? Are conversation texts sent to a third-party AI model, and where does it run? How long is data kept, and can it be deleted on request? Is the service ready to sign a data processing agreement?
Get your own side ready too. Clients should be told about the recording at the start of the call or in the terms on your website, and chats need a clear data processing policy. Keep the vendor's answers in writing: your lawyer and security team will need them.
What the price is made of
Speech analytics is priced in different ways, so comparing offers head-on is pointless. There are three models: paying per minute of calls, per manager seat per month, and per volume of analysed text. Transcription and meaning-level analysis are often billed separately and at different rates.
To compare suppliers, work out your own month. How many calls, how long they are, how many chats, how many managers. Ask each supplier to name a total for that volume, including recording storage, connecting telephony and the CRM, and setting up the checklist.
Ask about what the plan doesn't include. For example, a service charges modestly per minute, but every new checklist item is set up by its specialist for an extra fee. A month later you'll want to change the items, and the bill will grow where you didn't expect it.
Another question is plan limits. Are all calls reviewed or only some? Is there a cap on call length? Do you pay for short “couldn't get through” calls? If the service reviews a sample, you're back to manual spot checks, only now you're paying for them.
A pilot on your own calls
Decide on a service based on a pilot. A demo shows what the system can do in principle; a pilot shows what it does with your conversations. One or two managers with a live flow of calls and a couple of weeks of observation are usually enough.
- Gather a sample: twenty to thirty old calls and chats that the head of sales has already listened to and scored personally.
- Write three to five checklist items as plain sentences and give them to the service without hints.
- Compare the service's scores with your own for each conversation. Go through the discrepancies: where the system got it wrong and where you did.
- Connect the live flow: telephony, messengers, amoCRM. Check that each conversation lands in the deal card with no manual steps.
- A week later, look at the results: which deals the head of sales picked up from the report and how much time the review took.
The main pilot criterion is agreement with an experienced head of sales on old calls. If, in disputed cases, the system shows a quote and its conclusion can be checked in a minute, you can work with it. If the scores sound confident but the service can't back them up, there's no point in going on.
“Then our surveyor will come on Thursday at ten”listened to the end: the service was right
“The neighbours got it cheaper,” said the clientask the phone system for two-track recording
Hypothetical example. The main pilot criterion is agreement with an experienced sales head on past calls. Misses are counted only after review: some of them are the reviewer's own mistakes.
This scheme also works for checking our own AI Sales Control: it listens to calls and reads chats, checks them against a checklist, spots lost deals and scores managers. What the search for failed deals looks like is shown in the article about blown deals.
Mistakes when choosing a speech analytics service
The first mistake is choosing by the demo. A clean recording from a presentation says nothing about how the service will cope with a call from a noisy workshop. Testing on your own conversations takes a couple of days and removes the main risk.
The second is buying a keyword dictionary dressed up as analysis. Tags like “the word discount came up” look nice in a report but don't answer “did the manager book a measurement?” At the demo, ask how the system recognises a refusal that contains not a single word from the dictionary.
The third is reviewing calls separately from chats and the CRM. The conversation gets scored, yet nothing changes in the deal: no task is set, and the head of sales hears about the problem a week later. The fourth is starting with a thirty-point checklist. Managers can't tell what to fix first, and the review turns into a hunt for someone to blame.
The fifth is putting off the data question. A lawyer who learns after launch that recordings are being passed to a third-party service is entitled to stop the whole project. Ask the storage questions in the first week, while nothing is connected yet.
Speech analytics delivers more value together with the CRM, sales analytics and a robot that answers customers. How these pieces come together into one system is explained on our page on implementing AI in a business.
Frequently asked questions.
How is speech analytics different from ordinary call recording?
A recording stores sound, and to find out what was said you have to listen to it. Speech analytics turns the conversation into text and answers questions about it: was a next step set, did an objection come up, did the manager respond to it. The head of sales reads conclusions for every call instead of listening to a handful.
How many calls do you need for a speech analytics pilot?
To start with, twenty to thirty conversations that the head of sales has already scored personally are enough. They show whether the service's scores match yours and whether it can back up a conclusion with a quote. Then you connect the live flow of one or two managers and see whether the way deals are handled changes.
Can WhatsApp and Telegram chats be analysed together with calls?
Yes, if the service connects those channels. For a sales team this matters: clients often ask the key question in a chat and only then call. Check whether chats and calls are combined into one conversation per client. Then the score takes into account everything the manager has already written.
Is it legal to hand call recordings to an external service?
Yes, if you meet the requirements of the personal data law. Clients are warned about the recording, you sign a data processing agreement with the service, and data of Russian citizens is stored in databases located in Russia. Ask the vendor where its servers are and whether transcripts are passed to a third-party AI model.
Do I need amoCRM to connect speech analytics?
The service can work without a CRM, but it will be less useful. When the score and the conversation summary land in the deal card, the head of sales sees the problem where they actually work and can set the manager a task right away. Without a CRM the results sit in a separate dashboard that people visit less and less.
How much does a speech analytics service cost?
Pricing is per minute of calls, per manager seat or per volume of text, and transcription is often billed separately from the analysis. To compare, give suppliers your own month: number of calls, their length, chats and managers. Ask for a total that includes storage, connecting telephony and the CRM, and setting up the checklist.
Let's audit your business
Leave a request and on a free consultation we'll find where customers are being lost, suggest which service to start with and price it around your goal.
Rather not fill in a form? Message our agent on Telegram — it replies right away and books you a review.