Title: Voice to Text: How to Automate Calls and Correspondence
By 2026, the Kazakhstani market for voice technologies for business will reach 12 billion tenge — and demand continues to grow. According to local research, 67% of Kazakhs still prefer to resolve issues over the phone rather than through chat. Meanwhile, the global Voice AI market is estimated at $9.4 billion in 2025 and is growing at a rate of 28.5% per year. Modern speech recognition models achieve an accuracy of 95% and above on clean audio, and about 90% in real business negotiations. The volume of venture investments in AI startups in Kazakhstan increased from $14 million in 2023 to $73 million in 2025. These figures already make the technology practically applicable, not just demonstrative.
In our practice at West Star Ltd, we have implemented several projects where voice technologies addressed specific operational tasks: transcription of calls for subsequent analysis, automatic notifications to clients, preprocessing of incoming requests through a voice bot. In each case, the technology worked differently than initially expected — and it is this experience we want to share.
TWO DIRECTIONS — AND BOTH ARE IMPORTANT
When people say "voice to text," they mean automatic transcription: negotiations, calls, meetings, or voice messages are converted into text that can be analyzed, stored, and searched. The reverse direction is "text to voice," or speech synthesis: the system reads notifications, responses, instructions to clients.
In practice, both directions are often combined in one scenario. A client calls — the system recognizes his question (STT), refers to the knowledge base, forms a response, and reads it out loud (TTS). Or a manager held a meeting — the system transcribed the recording, highlighted key points, and added them to the CRM. In both cases, the automation chain begins with the machine "hearing" the person or "speaking" to him.
It is important to understand that these are not the same technology. STT and TTS are different tasks with different tools, different complexities, and different limitations. Mixing them into one "voice automation" is the first mistake that leads to an incorrect assessment of the project.
CALL TRANSCRIPTION: WHY DO IT
The most common application is the transcription of incoming and outgoing calls. According to our observations, most companies that record calls never listen to them: there is simply not enough time. Analyzing 500 calls manually takes about 41 hours of supervisor work. After implementing speech analytics, the same volume is sorted out in 15–20 minutes in report viewing mode.
Reading text is 5–8 times faster than listening to audio — this is a banal fact that becomes crucial when you have 200 calls a day. Transcription opens up several opportunities:
— content search (what the client said about delivery last week)
— quality control: did the manager follow the script, did he name the price correctly
— auto-filling the CRM card based on the conversation
— analytics: what objections are most common, what questions are repeated
— violation detection: foul language, deviation from regulations, unfulfilled promises
A McKinsey study notes that well-structured speech analytics increases customer satisfaction by 10% and reduces operating costs by 20–30%. Similar effects are confirmed by data from the Russian banking sector: one of the largest banks documented savings of 138 million rubles when deploying speech analytics for 5,000 employees. These figures are achievable, but not automatically — you need to understand what exactly to analyze and how to respond to it.
A separate story is voice messages in Telegram and WhatsApp. Business communication is increasingly moving to messengers, and voice messages have become the norm. The problem is that they are difficult to search, cannot be passed to a bot, and are difficult to include in formal document circulation. Automatic transcription of voice messages before bot processing is one of the most valuable and underrated scenarios: the client speaks with voice, the system translates it into text, the bot responds or routes.
SPEECH SYNTHESIS: WHERE IT IS NEEDED
The reverse task — making the system speak — is solved more easily and quickly. TTS technologies are now mature enough: voices sound natural, support Russian and Kazakh languages, and intonation can be adjusted. The delay in generating a short phrase is less than a second for most modern APIs.
Practical applications that really work:
— Voice notifications: "Your order is ready, expect the courier from 2 to 6 PM." Cheaper and faster than SMS, does not require a live operator.
— IVR menu: classic, but now the system can understand not only keystrokes but also spoken words — "say 'payment' or press 2."
— Voice bots for initial processing: the client names the order number or the essence of the request, the system routes the request.
— Reminders and confirmations: calling the database with reminders about meetings, collecting feedback with a short voice survey.
Here, the correct task setting is important. A voice bot that should "answer any questions" is a complex and expensive project. A voice bot that collects the order number and switches to the appropriate department is a realistic project for 2–3 weeks. The difference in complexity is fundamental.
TECHNOLOGIES: WHAT TO USE
For transcription in 2025–2026, a practical standard has formed. Whisper from OpenAI is one of the most accessible and accurate open-source models. Whisper supports 99 languages. On clean audio, according to the LibriSpeech test, it shows about 2.7% word error rate (WER). In real business negotiations, the figure rises to 8–12%. For the Russian language, there are fine-tuned versions: one of them reduced WER from ~9.8% to ~6.4% after fine-tuning on domain-specific Common Voice data.
In March 2025, OpenAI released GPT-4o-transcribe and GPT-4o-mini-transcribe — API solutions without public code, declared as more accurate successors to Whisper. There are no public benchmarks for the Russian language yet, but early tests show improvements on accented speech and noisy recordings.
For speech synthesis in Russian, several mature services are available: Russian (SaluteSpeech, Yandex SpeechKit) and international (Google TTS, Microsoft Azure TTS, ElevenLabs). The choice is determined by three factors: delay requirements (real-time or batch mode), cost per character or minute, and data localization issue — where the audio with client conversations goes.
A separate category is cloud platforms for comprehensive voice solutions: Deepgram, AssemblyAI, AWS Transcribe. They are convenient for a quick start, but as volumes grow, costs increase linearly. Self-hosted Whisper on your server or VPS is a more economical option for volumes of 5–10 hours of audio per day.
INTEGRATION INTO WORK PROCESSES
Technology itself is not a product. Value appears when recognized text gets to where people or other systems work with it.
Working integrations we observe in real projects:
— CRM (amoCRM, Bitrix24, 1C): automatic recording of call results, highlighting key agreements, setting tasks for the manager, tagging the client by the topic of the request.
— Messengers: transcription of voice messages in Telegram and WhatsApp before bot processing — the client speaks, the bot receives the text.
— HR and onboarding: voice input of data when filling out forms, voice applications for leave or business trips.
— Internal meetings: transcription, task highlighting with responsible persons, automatic protocol distribution to participants.
The most common mistake when implementing is to start with the most complex task. If a company wants "a full voice bot with understanding of arbitrary questions" right away, the project stretches for six months, the budget doubles, expectations are not met. It is better to start with the transcription of existing call recordings: this gives a quick result in 2–4 weeks and allows you to accumulate data for the next steps.
LIMITATIONS AND WEAKNESSES
An honest conversation about these technologies is impossible without listing the real problems.
— Accent and audio quality. Kazakh accent in Russian speech, background noise in the office, poor headset — all this sharply reduces recognition accuracy. The declared 95% is on clean studio audio. In a real call center, the figure is often 75–80%.
— Multilingualism in one conversation. Code-switching — switching from Russian to Kazakh in the middle of a sentence — is the norm in Kazakhstani business speech. Most models handle this poorly. Specialized solutions for Kazakh+Russian exist, but they are few.
— Industry-specific vocabulary. Specific terms, abbreviations, product names from a particular industry are often "heard" incorrectly by the model. Without fine-tuning or a user dictionary, accuracy in narrow niches drops.
— Latency in real-time recognition. Streaming transcription in real-time requires powerful servers or a cloud API with low latency. This is more expensive and complex than batch processing of already recorded audio.
— Legal and compliance risks. Call recording and transcription are regulated by law. Subscriber consent, data storage, cross-border transfer — if the API is located abroad — all this needs to be worked out before launch, not after.
— Cost at large volumes. Cloud APIs are charged per minute of audio. With tens of thousands of calls per month, the amounts become significant. Local deployment of Whisper is cheaper in the long run but requires infrastructure and DevOps resources.
PRACTICAL CONCLUSION FOR AUDIENCES
Specialist. Start with batch transcription of call recordings — this is the simplest entry into the topic. Whisper can be deployed locally, it is free and accurate enough for quality control tasks. For product integration, explore the GPT-4o-transcribe API — it has higher accuracy on complex audio, but there is a per-minute fee. Before choosing between cloud and self-hosted, calculate the actual volume of audio in minutes per month — scenarios diverge on this figure.
Manager. The main question is not "which technology to choose," but "which task are we solving." Transcription for manager control, a voice bot to unload the first line, notifications to clients — these are three different projects with different costs and risks. Do not agree to "do everything at once." A pilot on one scenario takes 4–8 weeks, a measurable result should be determined in advance: the percentage of calls covered by transcriptions, the time to process one request, the assessment of manager quality.
Owner. Speech technologies are not about replacing people, but about changing their workload. Supervisors listen to fewer calls manually and work more with anomalies and trends. Managers do not spend time writing meeting protocols. Operators answer fewer typical questions. This is an investment in data quality and managerial visibility, which pays off not immediately, but as processes are built around the information received.
FREQUENTLY ASKED QUESTIONS
Can calls be transcribed without the subscriber's consent?
No. Kazakhstani legislation requires notification of call recording. The standard practice is a voice warning at the beginning of the call. Using transcriptions within the company for analytics and training is permissible with the appropriate personal data processing policy.
How accurately is the Kazakh language recognized?
Significantly worse than Russian. Kazakh is a low-resource language for most commercial STT systems. Specialized solutions trained on Kazakh data exist, but they are few. Mixed speech (Kazakh+Russian) is a separate complex task for which there is currently no ready-made box solution of acceptable quality.
How much does it cost to implement call transcription for a small business?
A simple option: cloud API plus a script for integration with telephony. With a volume of 1,000 minutes per month, the API cost will be about $6. Development of integration — from 30 to 80 hours depending on your telephony. This is a realistic budget for a pilot, after which it is already clear whether to scale.
Is a voice bot the same as a chat bot?
No. A voice bot adds a layer of speech recognition on input and speech synthesis on output. The logic of request processing is the same as that of a chat bot, but recognition errors reduce the reliability of the entire chain. A voice bot is more complex to support: a text error is immediately visible, a voice error can lead to incorrect routing without an obvious error signal.