Speech-to-Text (STT / ASR)
in Today’s World
From HMM to Hybrid SALM Models; A Comprehensive Review of Architectures, Top 2025-2026 Models, Persian Challenges, and the Future of this Technology
What is STT and why is it important?
Speech-to-Text or ASR (Automatic Speech Recognition) is the technology that converts human spoken audio into written text. What seems simple is deceptive — behind every transcribed word lies decades of research in signal processing, linguistics, and deep learning.
The global market for this technology is on an unprecedented growth trajectory and will reach $53 billion by 2030. The primary driver of this growth is the large language model revolution and the emergence of hybrid LALM (Large Audio-Language Model) architectures that no longer merely transcribe — they “understand.”
Until 2022, ASR models “recognized patterns.” From 2023 onwards, new LALM models “understand” — they comprehend audio, not just transcribe it. This intrinsic difference in architecture, not just higher accuracy, is the key.
Evolution of STT Architectures
To understand where today’s models stand, we must know where we came from:
Hidden Markov Models with Gaussian Mixtures. The system required three independent components: acoustic model, pronunciation model, and language model. Low accuracy, reliance on experts, and fragility to noise were the main issues.
GMM · HMM · N-gram LMReplacing GMM with DNN was a massive leap. Later, RNN and LSTM entered for time-sequence modeling. These architectures still required a multi-stage pipeline.
DNN · LSTM · RNN · CTCConformer (combining Transformer + Convolution) changed everything. End-to-End models like wav2vec 2.0 and Whisper went directly from audio waveform to text. Trained on hundreds of thousands of hours of data with unprecedented performance.
Transformer · Conformer · wav2vec · WhisperHybrid SALM (Speech-Augmented Language Model) architecture. A robust audio encoder (like FastConformer) connected directly to an LLM. Models don’t just transcribe — they summarize and answer questions. NVIDIA Canary-Qwen was the first open-source example.
FastConformer · SALM · Multimodal ASREntirely new generation: Large Audio-Language Model. Models like Qwen3-ASR use Qwen3-Omni as a foundation and are fine-tuned with Reinforcement Learning. Result: exceptional resistance to noise, singing, accents, and complex textual patterns. Voxtral by Mistral also follows this path with a 4B streaming architecture.
LALM · Qwen3-Omni · AuT · GSPO-RL · Streaming 4BTop 2025 Models — Foundations of the New Generation
These models became the industry standard in 2025 and are still the right choice for many use cases:
First open-source SALM. FastConformer encoder + Qwen3-1.7B decoder. Capable of transcription, summarization, and answering questions. Led the Open ASR Leaderboard until late 2025 — now replaced by Qwen3-ASR.
Best throughput among all available models. Trained on 36,000 hours of audio (including silence and noise). For real-time captioning and call centers. Parakeet V3 (2026) released with 25-language support.
Ergodic Streaming Encoder for extremely low latency. Dynamic window instead of Whisper’s fixed 30-second chunk. Lower power consumption, ideal for mobile and IoT. The 6-person team rivaled NVIDIA and OpenAI.
Vast ecosystem, abundant documentation, easiest setup. Still the best choice for general multilingual projects. OpenAI also released GPT-4o-Transcribe and GPT-4o-mini-Transcribe models in March 2025.
New 2026 Models New
From late 2025 to early 2026, a new wave of models arrived that completely changed the competition. These are the models you need to know now:
Current leader of the Open ASR Leaderboard among open-source models. Trained on Qwen3-Omni foundation and fine-tuned with Reinforcement Learning (GSPO). Unique feature: can transcribe singing and music. 0.6B version for edge deployment, average TTFT 92ms. Accompanied by Qwen3-ForcedAligner-0.6B for precise timestamps.
Mistral’s 4B streaming model. Two versions: Mini Transcribe V2 (batch) and Realtime (streaming). Supports diarization, context biasing, and word-level timestamps. Claims to outperform GPT-4o mini Transcribe, Gemini 2.5 Flash, and Deepgram Nova — at nearly one-fifth the price of ElevenLabs Scribe.
The first model to simultaneously lead in AA-WER and latency. 2.3% WER on independent benchmarks — the lowest number available. The top choice for enterprise voice agents, though also the most expensive option.
Overall leader on Hugging Face Open ASR Leaderboard (including commercial). Built on Universal-2, which launched in August 2025 with 99 languages and built-in sentiment/entity/diarization. Unified Voice Agent API with STT+LLM+TTS on a single WebSocket.
Microsoft’s proprietary STT model. 3.0% AA-WER covering 140+ languages — best choice for Azure-embedded enterprises. Direct competitor to Google Chirp 3, which is also in public preview.
The cheapest high-quality option on the market. For high-volume batch transcription with limited budgets. Offers the best ROI if you only need supported languages and don’t require real-time processing.
Competition in 2026 is no longer just about WER. Deepgram Nova-3 with sub-300ms latency and Qwen3-ASR with proficiency across multiple languages and accents simultaneously demonstrate that the “best model” entirely depends on your use case.
Quantitative Comparison Table of Models (2025–2026)
The following numbers are based on the Hugging Face Open ASR Leaderboard, AA-WER v2.0 (Artificial Analysis), and official technical reports as of June 2026:
| Model | WER / AA-WER | RTFx / Latency | Language | Streaming | Type |
|---|---|---|---|---|---|
| ElevenLabs Scribe v2 2026 | ~150ms | Multilingual | Yes | Commercial | |
| Voxtral Small 2026 | <200ms | 13 Languages | Yes | Open-weight | |
| Azure MAI-Transcribe-1 2026 | ~140ms | 140+ | Yes | Commercial | |
| AssemblyAI Universal-3 Pro 2026 | ~150ms | 99+ | Yes | Commercial | |
| Qwen3-ASR-1.7B 2026 | 92ms (0.6B) | 52 Languages | Yes | Open (Apache 2) | |
| Canary-Qwen 2.5B | ×418 | English-centric | Limited | Open (Apache 2) | |
| Parakeet TDT 1.1B | >2000x | English | Yes | Open (Apache 2) | |
| Whisper Large V3 Turbo | ×6 vs V3 | 99+ | Partial | Open (MIT) | |
| Moonshine v2 | Very Low | English | Yes | Open (MIT) |
AA-WER v2.0 and standard WER are not the same. AA-WER is evaluated on more diverse data (VoxPopuli, Earnings-22, AMI-SDM, AgentTalk) and is typically higher than a model’s official WER. Always benchmark on your own data.
Why is Persian STT a Unique Problem?
Persian is one of the most challenging languages for speech processing. The issue isn’t just a lack of data — the linguistic structure of Persian is inherently more complex:
The Persian alphabet is right-to-left, and letters take different shapes in various positions. STT engines must account for this in text normalization.
From Tehrani to Isfahani, Shirazi, Mashhadi — each dialect has distinct pronunciations. Afghan Persian (Dari) and Tajik introduce additional challenges.
Persian has numerous compound words written either joined or separately: “Amir Reza” or “Amirreza” — a classic segmentation problem.
Persian is written with both Persian and Arabic numerals. The text contains many Arabic terms with varying pronunciations and spellings.
The Persian Common Voice is hundreds of times smaller than the English one. Many existing datasets have high noise — Google’s WER on Persian Common Voice is reported at 25.2%.
Most commercial applications of Persian STT deal with 8kHz telephone audio, which adds noise reduction challenges and increases error rates for general models.
Google STT’s WER on Persian Common Voice is 25.2%, and Whisper brings this number to 31.75%. These figures show that general models absolutely require dedicated fine-tuning for Persian.
Status of STT Models for Persian
The Persian STT ecosystem is growing. Major available models and approaches:
| Model / Approach | Type | Status | Key Note |
|---|---|---|---|
| Qwen3-ASR + Persian 2026 | Open-source + Custom | Under Review | Supports 52 languages — requires rigorous testing on conversational Persian. Its LALM architecture is promising for accents and noise. |
| Whisper + Fine-tuning Persian | Open-source + Custom | Most Common Approach | Fine-tuned on Persian Common Voice and proprietary datasets. The best balance between accuracy and execution speed for most use cases. |
| Google STT (Chirp 3) 2026 | Commercial API | Public Preview | The latest Chirp version with built-in denoiser and speaker diarization. Lowest cloud WER for Persian, but access restrictions in Iran persist. |
| wav2vec 2.0 + Persian | Open-source + Custom | Active in Research | Self-supervised on Persian audio without needing transcription — attractive for abundant unlabeled data. |
| Shenasa / HamTech ASR | Iranian | Limited | 30-hour open-source dataset. Good starting point for research, insufficient for production. |
| Domain-specific Conformer | Commercial/Custom | Best Result | Companies that have built custom models on domain-specific data (call centers, medical) achieve the best performance. |
Recommended Practical Approach for Persian
- Base: Test Whisper Large V3 Turbo or, if multilingualism is needed, Qwen3-ASR-1.7B
- Fine-tuning: At least 100 hours of labeled domain-specific audio for acceptable results
- Normalization: A robust post-processing layer is essential for homogenizing numbers, terms, and compound words
- Language Model: Integrate a Persian LM to improve output coherence — especially for proper nouns and specialized terms
- Forced Alignment: Use Qwen3-ForcedAligner-0.6B for precise timestamps — it outperforms WhisperX
Persian STT with sub-10% WER on real conversational audio (not benchmarks) is still a genuine gap in the market. The team that builds the best domain-specific Persian dataset — not necessarily the best model — will win this market.
The Future Path — Where Are We Going?
Qwen3-ASR and Voxtral showed that a strong foundation model + RL outperforms dedicated ASR architectures. By 2026-2027, every major player will have a proprietary LALM. Legacy multi-stage pipelines are no longer justifiable.
Nova-3 proved in 2025 that 10 languages can be processed simultaneously without routing. This is crucial for Persian — everyday spoken Persian is full of English, Arabic, and technical terms that the model must transcribe correctly.
Moonshine v2 with 245M parameters reached the accuracy of Whisper Large. Qwen3-ASR-0.6B boasts a 92ms TTFT. The path is clear: in the next 2 years, enterprise-quality STT will run on every smartphone — without the cloud.
Next-generation models will estimate not just text but also emotions, stress, speaker confidence, and even health status. Qwen3-ASR-Flash already demonstrated the ability to transcribe singing with background noise — paralinguistic analysis is the next step.
AssemblyAI, Azure, and Google are building interfaces allowing companies to fine-tune STT models on their own datasets without needing a machine learning team. This will be a massive differentiator for B2B service providers.
Industrial Applications of STT
Highest Overall Accuracy: ElevenLabs Scribe v2 · Best Multilingual Open-Source: Qwen3-ASR-1.7B · Fastest Throughput: Parakeet TDT · Cheapest Batch: Voxtral Mini · On-device: Moonshine v2 or Qwen3-ASR-0.6B · Enterprise Azure: MAI-Transcribe-1 · Specialized Persian: Whisper fine-tuned + Persian LM
Conclusion
From January to June 2026, the STT market saw four major shifts: Qwen3-ASR brought LALM+RL architecture to open-source, Voxtral demonstrated that 4B streaming could rival larger models, ElevenLabs Scribe v2 optimized WER and latency simultaneously for the first time, and Microsoft entered direct competition with MAI-Transcribe-1.
Three key takeaways for any team working in this field:
- Change Your Selection Criteria: WER is no longer the sole metric. AA-WER on diverse data, streaming latency, fine-tuning capability, and inference cost all matter in the final choice.
- For Persian, Data is More Important Than the Model: Qwen3-ASR supports 52 languages but still requires dedicated Persian fine-tuning. The team that builds the best domain-specific Persian dataset — not necessarily the best model — will win.
- LALM is the Future — Prepare Now: Legacy multi-stage pipelines are becoming obsolete. Investing in unified ASR+LLM architectures is justified starting now.
