AI Call Speech Analytics for Sales and Call Centers

Call speech analytics is the automatic transcription of recorded conversations and their evaluation against predefined criteria: whether the script was followed, whether the need was identified, whether objections were handled, whether timelines and price were stated, and whether there was a conflict. Its main purpose is coverage: instead of listening to a handful of calls, a manager sees an assessment of every conversation and reviews only the problematic ones. For implementation to deliver results, you need four things: lawful call recording with customer notification, speech recognition that handles telephone-quality audio, a clear scorecard with 5–10 criteria, and a check of AI scores against an experienced auditor’s scores before those scores are used for bonuses or penalties.

The material was prepared by AIrassvet based on open sources, verified on October 11, 2026. It is intended for business owners, sales directors, and heads of sales and contact centers who want to monitor call quality without manual listening. This is a practical guide: it includes solution selection, the legal minimum, a sample scorecard, and a method for checking accuracy. We did not test specific speech analytics services; the figures on recognition models are provided according to their developers. The section on call recording is not legal advice. AIrassvet is not affiliated with the companies mentioned, and the section at the end describes the publisher’s services.

What Call Speech Analytics Is
Quick Guide: How to Implement It Without Losing Data
What Speech Analytics Delivers: Coverage Instead of Sampling
How It Works: From Recording to CRM Scoring
Three Solution Types and How to Choose
Russian Speech Recognition: Open Models
Call Recording and the Law: Legal Minimum Requirements
Step-by-Step Implementation
Call Scorecard Template
How to Verify AI Scoring Accuracy
Which Metrics to Track
When Speech Analytics Is Not Needed
Common Mistakes
Frequently Asked Questions
Primary Sources and Verification Method

What Call Speech Analytics Is

Speech analytics has two parts. The first is speech recognition: the conversation recording is converted into text with separate turns for the sales rep and the customer. The second is text analysis: the system evaluates the conversation against criteria and assigns scores, tags, and a short summary.

In the past, analysis was rule-based: the system searched the transcript for keywords and phrases such as “good afternoon,” “promotion,” “too expensive,” or “competitor.” Now it is increasingly done by large language models: they understand the meaning of a turn, not just word matches. For example, they can distinguish “that’s expensive” as an objection from “it’s expensive if you don’t do it on time” as the manager’s argument.

A call transcript is not the goal; it is raw material. Value appears when call scores are linked to deals in the CRM and turned into decisions: who should be coached, which script needs to change, and which disputed case should be reviewed. For a broader look at sales quality control, including messaging, see the article on AI-based sales quality control.

Quick Guide: How to Implement It Without Losing Data

Start with the legal side: call recording notice at the beginning of the call, the purpose of processing, retention period, and access scope.
Check recognition quality on your own calls, not on the vendor’s demo: telephone audio, noise, and industry terminology matter.
Limit the scorecard to 5–10 criteria, each of which can be verified from the conversation text.
Compare AI scores with scores from an experienced auditor on 100–200 calls before launch.
Use the scores for coaching first, not for penalties, until accuracy is confirmed.
Link the scores to deals in the CRM, otherwise analytics will remain a report that nobody reads.

What Speech Analytics Delivers: Coverage Instead of Sampling

Manual quality control is almost always sample-based. Consider this example: a team has 10 sales reps, each making 60 calls per day, over 22 working days in a month. That is 13,200 calls per month. If an auditor listens to and evaluates 40 calls per day, they will cover 880 calls per month, or 6.7% of the total. The manager does not see the other 93% of conversations.

This is our example, not a statistic: actual numbers depend on call length and auditor workload. But the ratio is typical, and that is exactly what speech analytics changes. Automatic scoring covers every recorded call, while a human spends time reviewing the ones that received a low score or triggered an alert tag.

What this delivers in practice:

Script compliance across all calls, not just a random sample.
Finding problematic conversations: conflict, cancellation risk, competitor mention, complaint.
Training material: real examples of successful and unsuccessful objection handling. They can also be used in the AI training simulator for sales reps.
Feedback for marketing and product: what customers ask about and which objections come up repeatedly.
CRM updates: call summaries, agreements, and the next step are added to the deal record without manual entry. We described how to connect AI with a CRM in the article on the AI agent for Bitrix24 and amoCRM.

How It Works: From Recording to Scoring in CRM

A typical workflow looks like this:

Recording. Telephony or a virtual PBX records the conversation and attaches the recording to the call in the CRM. The Bitrix24 help center explicitly states that speech analytics apps require telephony with recording attachment enabled.
Recognition. The speech recognition model converts audio into text. Quality is measured using WER (word error rate) , the share of words misrecognized: the lower it is, the better.
Speaker separation. Turns are labeled as “sales rep” and “customer.” If the recording is dual-channel, this is easier; if it is single-channel, a separate separation model is needed, and errors there degrade the scores.
Analysis. Rule-based logic or a language model checks the conversation against the scorecard and generates tags, scores, and a summary.
Result. Scores and summaries are saved in the deal or call record, flow into reports, and alerts are sent to the manager for warning tags.

The weak link is usually transcription. A language model can’t fix what was misheard: if the text says “two thousand” instead of “two hundred thousand,” the call score will be wrong no matter how good the analysis is.

Three types of solutions and how to choose

Type	How it works	Strengths	Limitations	Best for
Telephony module or CRM app	A ready-made app pulls recordings from telephony and writes the result into the CRM	Fast start, no development required	Limited criteria customization, data is processed by the vendor	A small sales team using a popular CRM
Specialized speech analytics system	A separate platform with dictionaries, rules, and reports, often with LLM-based analysis	Deep reporting, customizable for a contact center	Cost and implementation time, separate interface	Contact center, large sales team
Your own in-house solution	An open speech recognition model and a language model on your company’s servers, integrated with telephony and CRM	Data stays in-house, full customization	Servers, integration, and maintenance are required	Banks, healthcare, companies with data storage requirements

The classification follows the market segmentation presented in the “Sam Reshu” review for July 2026: dictionary-based contact center systems, LLM analytics for sales teams, and built-in telephony modules. We combined the first two categories into specialized systems and added your own environment as a separate option.

How to choose:

If call volume is low, your CRM is popular, and data can be processed by the vendor — start with a CRM app. It’s the cheapest way to see whether the team will actually use the scores.
If you have a contact center with dozens of agents and complex scenarios — you need a specialized system with reports and rules. If you want to hand off some calls to a voice bot, also see the article on voice AI for call centers.
If the recordings contain medical, financial, or other sensitive data or security policy forbids external processing, choose your own environment.
If your scoring criteria are unique to your process — for example, checking technical order parameters — you’ll need LLM-based analysis with customizable instructions, either in a ready-made app or in your own solution.

We recommend starting with the simplest option that meets legal requirements, because the main implementation risk is not the technology, but that no one will use the scores. Trade-off: when switching to another solution, scorecards and integrations may need to be configured again.

Russian Speech Recognition: Open Models

For your own environment, you need a speech recognition model that can run on your own servers. Two well-known open models for Russian are T-one from T-Technologies and GigaAM from SberDevices.

According to the T-one repository, the model is distributed under the Apache 2.0 license, has 71 million parameters, processes audio in 300 ms chunks, and is designed for telephony. The developers provide the following WER comparison:

Dataset	T-one (71M)	GigaAM-RNNT v2 (243M)	Whisper large-v3 (1,540M)
Call center recordings	8.63%	10.22%	19.39%
Other telephony	6.20%	7.88%	17.29%
Names and titles	5.83%	9.55%	17.87%
CommonVoice 19 (read aloud)	5.32%	2.68%	5.78%

What this means:

The telephony-tuned model performs better on telephony. In this comparison, the general-purpose Whisper large-v3 makes mistakes in almost every fifth word in call center recordings.
There is no model that is best everywhere. On read-aloud speech, GigaAM-RNNT v2 has almost twice as few errors as T-one, so the choice depends on your recordings.
These are T-one developer data. Independent testing on your own calls is essential: accents, noise, industry terms, and product names change the result.

For a demo deployment, the T-one developers recommend at least 4 CPU cores and 8 GB of RAM; larger workloads require GPUs. The speech recognition model is only part of the solution: text analysis will also require a language model. We wrote about choosing one in the article on small language models and in the piece on local LLMs.

Call Recording and the Law: The Legal Minimum

Below is a summary of the requirements based on Kontur’smaterial, updated on May 21, 2026. This is not legal advice: for your specific situation, especially in healthcare, finance, and legal services, you should have a lawyer review it.

Conversations with customers. According to Kontur, there is no direct ban in Russia on recording conversations between managers and customers, but data from those recordings is processed under Personal Data Law No. 152-FZ. Minimum requirements:

notify the customer that the call is being recorded at the start of the call, usually with an automated message;
define the purpose of processing, for example, service quality control;
limit which employees have access to the recordings;
do not share recordings with third parties without the customer's consent;
set a retention period and delete the recordings after it expires.

Employee conversations. When recording for quality control, according to Kontur, formal employee consent is not required, but lawyers recommend including the recording in the employment contract, a local company policy, or a separate consent form.

When a voice becomes biometric data. If a recording is used to identify a person by voice, it becomes biometric personal data and written consent is required. Standard speech analytics for quality control does not identify a person by voice, but if a vendor offers “voice identification,” that is a different legal regime.

Consent and penalties. Starting September 1, 2025, consent for processing personal data can no longer be hidden in a contract or offer—it must be oformed as a separate document. Since May 30, 2025, the fine for legal entities for processing personal data without proper consent under Part 2 of Article 13.11 of the Russian Administrative Code is between 300,000 and 700,000 rubles.

Transferring recordings to an external service. If recognition and analysis are performed by a cloud service, this is a transfer of personal data to a processor, and if the server is abroad, it is also a cross-border transfer. More on this in the article on 152-FZ and foreign AI services. Kontur recommends servers located in Russia for AI services.

Retention period. The law does not set a single retention period. Kontur gives these guidelines: about 6 months for quality control, 1 year for disputed situations, and up to 3 years if the recording may be used as evidence in court.

Step-by-step implementation

Step 1. Define the question. Not “implement speech analytics,” but “find out why some inbound leads never make it to a meeting” or “check compliance with the initial consultation script.” The question determines the criteria and the solution you choose.

Step 2. Get your recordings in order. Make sure all calls are recorded, the recordings are attached to deals, and customers are notified about the recording. If the recordings have two channels—manager and customer separately—that will make analysis much easier.

Step 3. Build a control sample. 100–200 real calls from different managers, categories, and outcomes, including unsuccessful and conflict cases. For vendors, anonymize personal data in the sample.

Step 4. Check recognition accuracy. Transcribe 20–30 calls from the sample for each candidate and manually compare key points against the recording: amounts, deadlines, product names, and names. What matters is not average quality, but errors in the words that affect the evaluation.

Step 5. Create an evaluation scorecard. 5–10 criteria, each of which can be checked from the conversation text. A template is in the next section.

Step 6. Compare AI ratings with the reviewer’s ratings. The method is in the section on accuracy verification. If agreement on a criterion is low, rewrite the criterion or remove it.

Step 7. Connect CRM and reports. Put ratings, tags, and summaries into the deal record; send alert tags to the manager notification feed; and include a manager summary in the weekly report.

Step 8. Launch in training mode. During the first few weeks, the ratings are used for reviews and coaching, not for bonuses or penalties. Managers should see their ratings and have the ability to challenge them.

Step 9. Review the criteria monthly. Remove criteria that do not change anything, and add new ones based on review results.

Call evaluation scorecard template

Example for an initial call about an inbound lead. The criterion wording below is our example; adapt it to your own script.

No.	Criterion	How it is checked in the text	Response type	What to attach as evidence
1	Introduced themselves and named the company	Manager’s opening line within the first 30 seconds	Yes / no	Quote of the line
2	Identified the customer’s need	There are questions about the need and the customer’s response	Yes / partially / no	Manager’s questions
3	Clarified timing and budget	Timing or budget was stated, or the customer refused to state it	Yes / no	Quote
4	Handled the objection	If there was an objection, there is a substantive response	Yes / no / none	Objection and response
5	Did not make prohibited promises	No promises about discounts, timelines, or guarantees outside company policy	Violation / no	Quote of the promise
6	Set the next step	Agreed on a date, a meeting, or sending a proposal	Yes / no	Quote of the agreement
7	Mentioned a competitor	The customer named another company	Tag	Name and context
8	Conflict or Complaint	Strong wording, manager demand, threat to refuse	Tag, notification	Quote

Rules that make the scorecard workable:

Each criterion is checked against the text. “Was polite” is hard to verify, “did not interrupt the client more than three times” is verifiable if speaker separation works.
The model cites the quote. Require every score to be backed by a conversation excerpt: that way, a reviewer can verify a disputed rating in seconds.
There is a “not applicable” option. If there were no objections, the criterion “handled an objection” should not lower the score.
Emotions — as a tag only. Tone and sentiment scoring from text is unreliable; use it to find calls for review, not for bonuses.

How to Check AI Scoring Accuracy

Before launch, you need to understand how much you can trust AI scores. A repeatable process:

Take 100–200 calls from the control sample.
Have two experienced reviewers independently score them using the same scorecard. Discuss the calls where they disagreed and agree on a benchmark score — this will also show which criteria are phrased ambiguously.
Run those same calls through the system and compare the scores for each criterion separately.
Calculate three metrics for each criterion: the share of matches with the benchmark; false positives — the system flagged a violation that was not there; misses — a violation was there, but the system did not flag it.
Set the threshold in advance. For example, a criterion is allowed into bonus reports only if the match rate is no lower than the agreement rate between the two reviewers. Criteria below the threshold are used only as guidance for review.
Repeat the check after changing the model, script, or telephony provider.

Method limitation: a sample of a hundred or two calls will not reveal rare situations. So after launch, keep the option for managers to dispute a score and once a month manually review a small random sample.

What Metrics to Track

Speech analytics pays off if it changes business metrics, not just produces attractive reports. Choose 2–3 metrics before launch and record their current values:

Call-to-next-step conversion: meeting, proposal sent, payment.
Share of calls with violations of key criteria — by manager and by week.
Reviewer time per analyzed call — should go down, because the system highlights the problem area.
Response time to complaints and threats to refuse — from the call to the manager’s action.
CRM completeness after calls, if the system writes summaries into the deal card.

Do not expect an immediate effect from analytics alone: results appear when the data is used to change scripts and provide training. Vendor promises of a specific percentage increase in conversion are their own case studies and cannot be applied directly to your process.

When Speech Analytics Is Not Needed

There are too few calls. If the team has only a dozen conversations a day, the manager can listen to them personally, and the time is better spent on the script and training.
Sales do not happen by phone. If the main negotiations take place in messengers and email, you need to analyze written communication — a similar but different task.
The recordings are not linked to deals. Without a “call — deal — result” connection, analytics will show violations, but not which ones affect sales.
There is no one to review the results. Scores without reviews and training change nothing.
It is not legally allowed to record and process calls in the chosen setup — address the legal issue first.

Common Mistakes

Too many criteria. A 30-point scorecard does not fit in the head of either the manager or the supervisor.
Trusting scores without cross-checking. If the system gives low scores for things that did not happen and nobody checks it, the team quickly stops trusting it.
Penalties from day one. Managers start “talking for the robot,” not for the customer.
Testing on demo recordings. Studio-quality audio and standard phrases do not show how the system will handle your calls.
Forgetting the legal side. No call-recording notice or indefinite storage of recordings creates risks that outweigh the benefits of analytics.
Reports that nobody opens. Results need to appear where managers actually work: in CRM, notifications, and weekly meetings.

Frequently Asked Questions

What is call speech analytics?

It is automatic transcription of recorded phone conversations and evaluation against set criteria: script compliance, needs discovery, objection handling, commitments, conflicts. The results are saved in CRM and reports, and problematic calls are sent to the manager for review.

Is it legal to record customer calls for analysis?

According to Kontur, there is no direct ban on recording conversations between managers and customers, but you must notify the customer about the recording, define the purpose, restrict access, set a retention period, and not share recordings with third parties without consent. For sensitive industries, written consent and legal review are recommended.

Is the customer's voice biometric data?

Only if the voice is used to identify the person, for example for voice identification. In that case, written consent is required. According to Kontur, recording a conversation for quality control is not considered biometric data.

How is LLM-based analytics better than keyword-based analytics?

A keyword-based system looks for specific words and phrases, while a language model understands the meaning of an utterance and can check criteria that are not reduced to keywords, such as whether a specific next step was set. But LLM analysis is more expensive and also makes mistakes, so it needs to be cross-checked against a reviewer’s scoring.

Can speech analytics be deployed on your own servers?

Yes. For speech recognition, there are open models for Russian, such as T-one and GigaAM, and for analysis there are language models that can run locally. You will need servers, integration with telephony and CRM, and ongoing support for the solution.

How many calls are needed to test the system before launch?

We recommend 100–200 real calls from different managers and outcomes, scored by two experienced reviewers. That is enough to spot weak criteria and recognition errors, but not enough for rare situations, so sample checking is still needed after launch.

Primary Sources and the Verification Method

We reviewed the primary sources on October 11, 2026. Data on the T-one model and the WER comparison are taken from the developer’s repository; call-recording requirements are from Kontur material; operating conditions for speech analytics apps are from Bitrix24 help; market classification is from the Sam Reshu review. The implementation steps, scorecard template, accuracy-check method, and coverage calculation are AIrassvet’s editorial methodology, not the result of testing specific services.

“T-Technologies:” T-one repository on GitHub — license, model size, WER comparison, deployment requirements.
“Kontur,” 05/21/2026: recording phone calls with customers and employees — notice, consent, retention, biometrics, fines.
Bitrix24, 08/28/2026: speech analytics for calls — apps — connection terms and app criteria.
“Sam Reshu,” July 2026: speech analytics in 2026 — solution categories on the market.

Limitations: the model comparison was conducted by the T-one developer and has not been independently verified; legal requirements are paraphrased from a secondary source and may differ for industries with special regulation; the example of control coverage is based on illustrative data.

How AIrassvet helps implement speech analytics for calls

If you want to analyze calls within your own environment or tie conversation scoring to deals in your CRM, it is better to build the solution around your sales process rather than around a ready-made set of reports. On the AIrassvet website you’ll find the directions that relate directly to this task:

Integrating multimodal AI models — speech recognition and a language model — into company workflows.
Deploying language models on company servers when call recordings cannot be sent to an external service.
AI agents integrated with communication channels, including telephony, and the company’s business systems, such as CRM.
Process audit, data preparation, testing, deployment, and ongoing support for the solution.

The first step is to choose one type of call, describe your current metrics, where the recordings are stored, what data restrictions apply, and what criterion you will use to consider the system’s scores accurate enough.

Discuss the project.

Bottom line: accuracy and compliance first, reports second

Speech analytics for calls shifts quality control from samples to all conversations, but on its own it does not improve anything. Results come from four conditions: lawful recording with customer notice and a clear retention period, recognition tested on your own call recordings, a concise scoring rubric with evidence quotes, and a comparison of AI scores with those of an experienced quality controller before those scores are used to make decisions about people.

Start with one type of call and a control sample of 100–200 conversations. If the system matches the controller on the key criteria, connect CRM and reporting; if not, rewrite the criteria or change the recognition model before scaling.