tyger-ears-turkish

Barge-in and end-of-turn detection for Turkish telephone voice agents.

This is TYGER by NineNineSix with its turn-taking controller fine-tuned for Turkish phone calls answered by a text-to-speech agent. The model listens to two audio streams, the caller's microphone and the agent's own playback, and reports two events:

  • barge-in: the caller started talking over the agent, so the agent should stop speaking;
  • endpoint: the caller finished their turn, so the agent can answer.
Input two aligned mono 16 kHz streams: the caller's microphone and the agent's playback
Output barge-in and endpoint events, decided every 80 ms on audio fed in 320 ms hops
Size 652M parameters in FP32 (2.61 GB): NVIDIA Nemotron ASR (638M, unchanged) and a GRU controller (14M)
Hardware an NVIDIA GPU, about 3 GB of memory
License OpenMDW-1.1

What changed from TYGER

Only the controller changed: its 39 tensors (turn_taker.*) were fine-tuned on Turkish two-channel telephone material. The Nemotron weights (nemotron.*) are byte-identical to the original release.

The original controller was trained on English conversations between people. Over a synthetic agent voice it missed most interruptions, and it ended Turkish turns early. Results on our held-out test (real Turkish callers over a text-to-speech agent; the test audio is not public):

TYGER this model
Interruptions detected, of 30 5 to 7 30
False barge-ins on the agent's own echo, per minute not measured 0
End of turn: right / early / missed, of 164 123 / 38 / 3 158 / 3 / 3
Delay from the caller stopping to the endpoint, median 0.36 s 0.32 s

A barge-in comes a median 0.74 s after the caller starts speaking.

How to use

The model runs with the open-source TYGER runtime, tyger-s2s (Apache-2.0), on an NVIDIA GPU.

1. Get the runtime and the model

git clone https://github.com/nineninesix-ai/tyger-s2s
cd tyger-s2s          # install its requirements as its README describes
hf download wite-tech/tyger-ears-turkish --local-dir tyger-ears-turkish

2. Feed it audio

Run this from the tyger-s2s folder. Open one session per call. Every 320 ms, pass the caller's microphone and what the agent played in the same 320 ms.

import asyncio
import numpy as np
from app.plugins.ears.tyger.engine import EarsEngine
from app.plugins.ears.tyger.service import EarsService

HOP = 5120  # 320 ms at 16 kHz


async def main():
    service = EarsService(EarsEngine("tyger-ears-turkish", slots=4))  # slots: calls at once
    session = service.open(ep_thr=0.91, barge_thr=0.7, bc_thr=2.0)    # one session per call

    # float32 arrays of 5,120 samples in [-1, 1]; silence here, your real audio in an application
    microphone = np.zeros(HOP, np.float32)
    playback = np.zeros(HOP, np.float32)
    mode = "listen"  # "speaking" while the agent's audio is playing, "listen" otherwise

    for event in await service.feed(session, microphone, playback, mode):
        if event["kind"] == "BARGE":
            pass  # stop the agent's audio and switch to mode "listen"
        elif event["kind"] == "ENDPOINT":
            pass  # the caller finished: transcribe the turn and answer

    service.close(session)


asyncio.run(main())

To try it inside the tyger-s2s voice agent instead, set ears.model in its config/default.yaml to the downloaded folder.

Settings we use: barge_thr=0.7 and ep_thr=0.91, each over two consecutive frames (the runtime's default). bc_thr=2.0 switches the backchannel events off.

For good results

  • Keep both streams on one clock. The playback stream must be what the caller actually hears, within about 0.4 s of its echo in the microphone.
  • Use mode "speaking" only while the agent's audio is really playing.
  • Add your own silence timeout. The runtime ends a turn only when the model says so; we also end a turn after 1.2 s of silence and cap it at 20 s.
  • Transcribe the caller's turn with your own recogniser. The Nemotron recogniser in the bundle is NVIDIA's original and was not changed or evaluated here.

Training

  • Starting point: TYGER (nineninesix/tyger, revision cd826db). All controller weights were fine-tuned; Nemotron stayed frozen.
  • Data: about 11,000 synthetic two-channel cases. The agent channel is text-to-speech sentences in four voices. The microphone channel is the agent's echo at varying levels and delays, partly through a room response and a narrowband telephone line, plus Turkish telephone caller speech and background noise.
  • Labels: by construction. Barge-in while the caller speaks over the agent; endpoint from the end of the caller's utterance.

The caller recordings are proprietary and are not released. This repository contains no audio.

Limitations

  • Trained and tested on synthetic two-channel mixes and on single-channel real clips, with four text-to-speech voices as the agent. Another agent voice may need new fine-tuning: a voice missing from training caused false barge-ins on the agent's own echo.
  • Built for Turkish telephone speech. An English control test still passed; nothing else was measured.
  • A playback stream that is 0.8 s or more out of line with its echo causes missed and false barge-ins.
  • Very quiet callers are detected less often: 25 to 27 of 30 interruptions at -35 dBFS and below, against 29 to 30 at normal levels (measured without echo in the microphone).

License

The whole repository is released under OpenMDW-1.1: see LICENSE.md and NOTICE.md.

  • The controller is derived from TYGER, copyright 2026 NineNineSix.
  • Nemotron 3.5 ASR Streaming 0.6B is NVIDIA's, under NVIDIA's own OpenMDW-1.1 grant. Its model card, figures and license text are unchanged in upstream/ and licenses/.

The runtime code is not part of this repository and keeps its own license (Apache-2.0).

Citation

Please cite the model this one is built on:

@misc{tyger_2026,
  author    = {{NineNineSix}},
  title     = {{TYGER}: Speech Recognition and Duplex Turn Control},
  year      = {2026},
  publisher = {NineNineSix},
  url       = {https://huggingface.co/nineninesix/tyger}
}
Downloads last month
2
Safetensors
Model size
0.7B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for wite-tech/tyger-ears-turkish

Finetuned
(1)
this model

Collection including wite-tech/tyger-ears-turkish