How to connect AVA with T-one (russian language best choice for STT)

[dowletos]

Dear all appreciate your assistance on how to connect AVA with T-one. I build a full local and want to change Vost russian STT to T-one STT. which provides around 95% accuracy in russian language speech recognition.
Appreciate your help

[linear]

AAVA-203

[dowletos]

Here is T-one STT model description:
https://github.com/voicekit-team/T-one

[hkjarral]

Hi @dowletos,

Thank you for your interest in integrating T-one STT! This is a great choice for Russian telephony applications - T-one is specifically optimized for 8kHz telephony audio which aligns well with AVA's audio pipeline.

Integration Options

You have two paths to integrate T-one:

Option 1: Local AI Server Backend (Recommended for Full Local)

Since you're running a full local setup, the cleanest approach is adding T-one as a new STT backend in local_ai_server/backends/stt/.

Step 1: Create the backend file

Create local_ai_server/backends/stt/tone_backend.py following the pattern in vosk_backend.py:

from __future__ import annotations
from typing import Any, Dict, Optional
from backends.interface import STTBackendInterface

class TOneBackend(STTBackendInterface):
    def __init__(self):
        self._pipeline = None

    @classmethod
    def name(cls) -> str:
        return "tone"  # This becomes the config key

    @classmethod
    def config_schema(cls) -> Dict[str, Any]:
        return {
            "model_path": {
                "type": "string",
                "required": False,
                "default": "t-tech/T-one",
                "description": "Path to T-one model or HuggingFace repo"
            },
            "device": {
                "type": "string",
                "required": False,
                "default": "cpu",
                "enum": ["cpu", "cuda"]
            },
        }

    @classmethod
    def is_available(cls) -> bool:
        try:
            import onnxruntime
            return True
        except ImportError:
            return False

    def initialize(self, config: Dict[str, Any]) -> None:
        # Load T-one pipeline here
        # See: https://github.com/voicekit-team/T-one
        pass

    def shutdown(self) -> None:
        self._pipeline = None

    def process_audio(self, audio_bytes: bytes) -> Optional[str]:
        """
        Process audio and return transcript.
        
        T-one expects 8kHz 16-bit PCM audio in 300ms chunks.
        AVA sends audio at 16kHz by default, so you may need to resample.
        """
        if not self._pipeline:
            return None
        # Your T-one integration here
        return None

    def status(self) -> Dict[str, Any]:
        return {"backend": "tone", "loaded": self._pipeline is not None}

Step 2: Register the backend

Add to local_ai_server/backends/stt/__init__.py:

from backends.stt.tone_backend import TOneBackend
STT_REGISTRY.register(TOneBackend)

Step 3: Configure in YAML

local_ai_server:
  stt:
    backend: tone
    model_path: "t-tech/T-one"
    device: cuda  # or cpu

Option 2: Pipeline Adapter (For WebSocket/REST Endpoint)

If you prefer running T-one as a separate service (e.g., using the t-one-rest-api wrapper), create a pipeline adapter in src/pipelines/tone.py.

See src/pipelines/deepgram.py for a complete example of:

• REST API mode (transcribe() method)
• WebSocket streaming mode (start_stream(), send_audio(), iter_results())

Key Technical Details

T-one Audio Requirements:

• Sample rate: 8kHz (telephony-optimized)
• Format: 16-bit PCM
• Chunk size: 300ms (~4800 bytes per chunk)

AVA Audio Pipeline:

• Default local STT format: 16kHz PCM16
• You'll need to downsample from 16kHz → 8kHz
• Use src/audio/resampler.py utilities:

  from src.audio.resampler import resample_audio
  audio_8k, state = resample_audio(audio_16k, 16000, 8000, state=None)

Development Workflow

1. Fork the repo and create a feature branch
2. Use AI-assisted coding: Open in Windsurf/Cursor with our IDE rules
3. Follow our guides:
   • Provider Development
   • Architecture Quickstart
   • Common Pitfalls
4. Test with a real call and collect logs via ./scripts/rca_collect.sh
5. Submit a PR - we'd love to see T-one as an officially supported backend!

Resources

Resource	Link
T-one Model	t-tech/T-one on HuggingFace
T-one Repo	voicekit-team/T-one
REST API Wrapper	masasibata/t-one-rest-api
Backend Interface	local_ai_server/backends/interface.py

Feel free to reach out on Discord if you need help during implementation. This would be a valuable contribution to the Russian-language telephony community! 🇷🇺

[dowletos]

Dear @hkjarral
Thank you for your comprehansive response.
I found this speech recognition stt tools testing feature (https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition)
And you know what:

This STT:
"sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16"
Decoding method: greedy_search
Whether to add punctuation (Only for Chinese): Yes
This STT provides amazing performance 1 minutes speach recoring takes 0.4 sec only for convertion to text and it has recognized the long long texts without even 1 error. 100% correct recognition. Amazing.
The question is who could help me to integrate this to the AVA.

And sherpa has all ready build including T-one and GigAM (better than T-one) availiable for download and integration.

Could you be so kind to assist me with integration. Once done it would be ready to go full offline build for russian language streaming recognition.
Here is the list of all other sherpa builds availiable: https://k2-fsa.github.io/sherpa/onnx/index.html
Would appreciate so much. Ready to do anyting would be required.
Thank you @hkjarral in advance.

[dowletos]

Could someone help me with integration GigaAM STT to AVA?
Exactly this model provides the best productivity and accuracy. amazing how quick it is:
Name: v3_e2e_rnnt "sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16.tar.bz2"
https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16.tar.bz2

Bytheway here you could test productivity and accurasy of any existing models 👍
https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition

The probles is I just don't know how to integrate it to AVA.
I have already installed Sherpa_English language STT availiable at AVA and runned it . All is ok. But when i switched the Sherpa_English to the russian model downloaded i got an error loadint STT to AVA:

2026-03-03 12:10:58 +0000 DEBUG asyncio: Using selector: EpollSelector
2026-03-03 12:10:58 +0000 INFO root: 🚀 Initializing enhanced AI models for MVP...
2026-03-03 12:10:58 +0000 INFO root: 🎤 STT backend: Sherpa-onnx (local streaming ASR)
2026-03-03 12:10:58 +0000 INFO root: 📁 SHERPA - Model files found:
2026-03-03 12:10:58 +0000 INFO root: tokens: /app/models/stt/sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16/tokens.txt
2026-03-03 12:10:58 +0000 INFO root: encoder: /app/models/stt/sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16/encoder.int8.onnx
2026-03-03 12:10:58 +0000 INFO root: decoder: /app/models/stt/sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16/decoder.onnx
2026-03-03 12:10:58 +0000 INFO root: joiner: /app/models/stt/sherpa-onnx-nemo-transducer-punct-giga-am-v3-russian-2025-12-16/joiner.onnx

/project/sherpa-onnx/csrc/online-transducer-nemo-model.cc:InitEncoder:323 'window_size' does not exist in the metadata

don't know hot to fix it

[hkjarral]

I ll explore this further on this branch, i ll share instructions on how to test and verify it

https://github.com/hkjarral/AVA-AI-Voice-Agent-for-Asterisk/tree/russian-stt-support

[dowletos]

Thank you for your contribution!

[dowletos]

I ll explore this further on this branch, i ll share instructions on how to test and verify it

https://github.com/hkjarral/AVA-AI-Voice-Agent-for-Asterisk/tree/russian-stt-support

Your support is awesome!!!

[dowletos]

I ll explore this further on this branch, i ll share instructions on how to test and verify it

https://github.com/hkjarral/AVA-AI-Voice-Agent-for-Asterisk/tree/russian-stt-support

Eventhough, I would so much appreciate your direct help with integrating this GigaAM model (https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models) directly to AVA.
Me personally not a software engineer and it could take months till i will get into the AVA project deeply. Appreciate your help it would help a thousands of people. thank you again.

[hkjarral]

@dowletos have you tested faster whisper ? How does it perform with russian language ?

If not please try it

Set FASTER_WHISPER_LANGUAGE=ru in .env file in the project directory.

Recreate ai engine and test it.

[dowletos]

@dowletos have you tested faster whisper ? How does it perform with russian language ?

If not please try it

Set FASTER_WHISPER_LANGUAGE=ru in .env file in the project directory.

Recreate ai engine and test it.

Will carry out tests today and will write a feedback asap.

[hkjarral]

@dowletos Any luck with the testing ?

[dowletos]

Hello Hkjarral. Sorry set evrything up but some issue raised in other production systems. Was forced to stop testings. Tomorrow will make you know the resulta. Thank you again.

…

On Sat, 7 Mar 2026 at 01:04 Haider Jarral ***@***.***> wrote: *hkjarral* left a comment (hkjarral/AVA-AI-Voice-Agent-for-Asterisk#258) <#258 (comment)> @dowletos <https://github.com/dowletos> Any luck with the testing ? — Reply to this email directly, view it on GitHub <#258 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AG3ZC5ZPXLFX7LQRB6O562D4PMVMJAVCNFSM6AAAAACWD7SXO2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DAMJTHA2TMNRSGE> . You are receiving this because you were mentioned.Message ID: ***@***.*** com>