From 15b6a00c7c2465f6a54e15e06678fab6f2c676d5 Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Wed, 29 Apr 2026 07:30:21 +0000
Subject: [PATCH 01/13] Add canary streamatt

---
 config/canary_streamaat.yaml                  |  13 ++
 .../speech_processors/canary_streamatt.py     | 167 ++++++++++++++++++
 2 files changed, 180 insertions(+)
 create mode 100755 config/canary_streamaat.yaml
 create mode 100755 simulstream/server/speech_processors/canary_streamatt.py

diff --git a/config/canary_streamaat.yaml b/config/canary_streamaat.yaml
new file mode 100755
index 0000000..861a4f7
--- /dev/null
+++ b/config/canary_streamaat.yaml
@@ -0,0 +1,13 @@
+type: "simulstream.server.speech_processors.canary_streamatt.CanaryStreamAtt"
+model_name: "nvidia/canary-1b-v2"
+text_history:
+  type: "simulstream.server.speech_processors.base_streamatt.FixedWordsTextHistory"
+  history_words: 10
+speech_chunk_size: 0.960  # seconds
+detokenizer_type: "canary"
+cross_attn_layer: -2
+cutoff_frame_num: 8
+num_beams: 5
+audio_history_max_duration: 160  # Maximum length for the audio buffer, in seconds
+text_history_max_len: 128
+word_level_postprocess: True  # Disable if character-level language
diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
new file mode 100755
index 0000000..cec237c
--- /dev/null
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -0,0 +1,167 @@
+# Copyright 2025 FBK
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+
+# http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License
+
+import logging
+import torch
+import numpy as np
+
+from types import SimpleNamespace
+from typing import List, Tuple
+
+from dataclasses import replace
+import copy
+
+from simulstream.server.speech_processors import SAMPLE_RATE
+from simulstream.server.speech_processors.base_streamatt import BaseStreamAtt, BOW_PREFIX
+
+from nemo.collections.asr.models import ASRModel
+from nemo.collections.asr.parts.submodules.multitask_decoding import (
+    MultiTaskDecodingConfig,
+)
+from nemo.collections.asr.models.aed_multitask_models import (
+    MultiTaskTranscriptionConfig,
+)
+
+logger = logging.getLogger(__name__)
+
+MEL_HOP_SAMPLES = 160
+CANARY_AUDIO_SUBSAMPLING = 8
+
+
+class CanaryStreamAtt(BaseStreamAtt):
+    """
+    StreamAtt policy implementation for NVIDIA's Canary-v2 model
+
+    Args:
+        config (SimpleNamespace): Configuration object.  
+            Supported attributes:
+            - **pnc (str)**: ``"yes"`` for punctuation/capitalisation, ``"no"`` otherwise.
+              Defaults to ``"yes"``.
+            - **audio_history_max_duration (int)**: Maximum audio history in seconds.
+              Defaults to ``30``.
+    """
+
+    def __init__(self, config: SimpleNamespace):
+        super().__init__(config)
+        self.use_raw_audio_history = True
+        self.mel_hop_samples = MEL_HOP_SAMPLES
+        self.audio_subsampling_factor = CANARY_AUDIO_SUBSAMPLING
+        self._pnc: str = getattr(self.config, "pnc", "yes")
+        self._audio_history_max_duration: int = getattr(self.config, "audio_history_max_duration", 30)
+
+        # Build the transcription config, which is reused for every transcribe() call.
+        self.transcription_cfg = MultiTaskTranscriptionConfig(
+            batch_size=1,
+            return_hypotheses=True,
+            enable_chunking=False,
+            verbose=False,
+        )
+
+    @property
+    def audio_max_len(self) -> int:
+        """Maximum audio history length in raw waveform samples."""
+        return self._audio_history_max_duration * SAMPLE_RATE
+
+    def set_source_language(self, language: str) -> None:
+        self.src_lang = language
+
+    def set_target_language(self, language: str) -> None:
+        self.tgt_lang = language
+
+    @classmethod
+    def load_model(cls, config: SimpleNamespace):
+        if not hasattr(cls, "model") or cls.model is None:
+            cls.model = ASRModel.from_pretrained(model_name=config.model_name)
+
+            # Configure decoding strategy
+            multitask_decoding = MultiTaskDecodingConfig()
+            multitask_decoding.strategy = "beam"
+            multitask_decoding.return_xattn_scores = True
+            multitask_decoding.beam.beam_size = getattr(config, "num_beams", 5)
+            cls.model.change_decoding_strategy(multitask_decoding)
+
+            cls.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+            assert cls.model.cfg.preprocessor.sample_rate == SAMPLE_RATE
+            cls.model.to(cls.device)
+
+    def _build_transcription_config(self):
+        """
+        Return a ``MultiTaskTranscriptionConfig`` whose prompt encodes the current source/target
+        languages, task, PNC preference, and forced decoder prefix.
+        """
+
+        default_turns = self.model.prompt.get_default_dialog_slots()
+        default_slots = copy.deepcopy(default_turns[0]["slots"])
+        default_slots["source_lang"] = self.src_lang
+        default_slots["target_lang"] = self.tgt_lang
+        default_slots["pnc"] = self._pnc
+
+        turns = [
+            {
+                "role": "user", "slots": default_slots
+            },
+            {
+                "role": "user_prefix",
+                "slots": {
+                    "prefix": self.model.tokenizer.tokens_to_text(self.text_history)
+                },
+            },
+        ]
+
+        return replace(self.transcription_cfg, prompt={"turns": turns})
+
+    def _remove_eos_tokens(self, token_ids: List[int]) -> List[int]:
+        """Strip leading EOS tokens that the model may prepend when a forced prefix is used."""
+        if not token_ids:
+            return token_ids
+        pos = 0
+        while pos < len(token_ids) and token_ids[pos] == self.model.tokenizer.eos_id:
+            pos += 1
+        return token_ids[pos:]
+
+    def _preprocess(self, waveform: np.ndarray) -> np.ndarray:
+        """
+        Append the incoming waveform chunk to the raw audio history and return it.
+
+        Returns:
+            np.ndarray: Accumulated raw audio history.
+        """
+        if self.audio_history is None:
+            self.audio_history = waveform.astype(np.float32)
+        else:
+            self.audio_history = np.concatenate(
+                [self.audio_history, waveform.astype(np.float32)])
+
+        return self.audio_history
+
+    def _generate(self, speech: np.ndarray) -> Tuple[List[str], torch.Tensor]:
+        override_config = self._build_transcription_config()
+
+        with torch.inference_mode():
+            output = self.model.transcribe(audio=speech, override_config=override_config)
+
+        hypothesis = output[0]
+
+        token_ids = hypothesis.y_sequence.detach().cpu().tolist()
+        token_ids = self._remove_eos_tokens(token_ids)
+        tokens: List[str] = self.model.tokenizer.ids_to_tokens(token_ids)
+
+        xatt_raw = hypothesis.xatt_scores[self.cross_attn_layer]
+        xatt = xatt_raw.mean(dim=0).cpu()
+        xatt = self.normalize_attn(xatt)
+
+        return tokens, xatt
+
+    def tokens_to_string(self, tokens: List[str]) -> str:
+        return self.model.tokenizer.tokens_to_text(tokens)
\ No newline at end of file

From 4279681f7f6bbddd989e25ae7708cb97c6a58281 Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Wed, 29 Apr 2026 07:31:28 +0000
Subject: [PATCH 02/13] Add audio history type flag to the base streamatt

---
 simulstream/server/speech_processors/base_streamatt.py | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/simulstream/server/speech_processors/base_streamatt.py b/simulstream/server/speech_processors/base_streamatt.py
index fa7ccd4..b0899e3 100644
--- a/simulstream/server/speech_processors/base_streamatt.py
+++ b/simulstream/server/speech_processors/base_streamatt.py
@@ -60,6 +60,10 @@ class BaseStreamAtt(BaseSpeechProcessor):
                  context for next predictions.
            - **audio_subsampling_factor (int)**: Subsampling factor of the model, if any.
              Defaults to 1.
+           - **mel_hop_samples (int)**: Number of raw waveform samples per mel frame.
+             Defaults to 1.
+           - **use_raw_audio_history (bool)**: Returns whether ``audio_history`` stores raw waveform samples
+             rather than processed frames. Defaults to False.
            - **text_history_max_len (int)**: The maximum length of the textual history after which
              the current content is cut. Defaults to 128.
            - **cross_attention_layer (int)**: Layer from which to extract the cross-attention from.
@@ -77,6 +81,8 @@ def __init__(self, config: SimpleNamespace):
         text_history_cls = class_load(text_history_config.type)
         self.text_history_method = text_history_cls(text_history_config)
         self.audio_subsampling_factor = getattr(self.config, "audio_subsampling_factor", 1)
+        self.mel_hop_samples = getattr(self.config, "mel_hop_samples", 1)
+        self.use_raw_audio_history = getattr(self.config, "use_raw_audio_history", False)
         self.text_history_max_len = getattr(self.config, "text_history_max_len", 128)
         self.cross_attn_layer = getattr(self.config, "cross_attention_layer", 3)
         self.cutoff_frame_num = getattr(self.config, "cutoff_frame_num", 2)
@@ -176,6 +182,10 @@ def _update_speech_history(self, discarded_text: int, cross_attn: torch.Tensor)
         # Multiply by the subsampling factor to recover the original number of frames
         frames_to_cut = earliest_attended_idx * self.audio_subsampling_factor
 
+        # If audio is stored as raw waveform, convert frames to cut into number of samples to cut
+        if self.use_raw_audio_history:
+            frames_to_cut = frames_to_cut * self.mel_hop_samples
+
         # Cut the unattended audio features
         self.audio_history = self.audio_history[frames_to_cut:]
 

From 53afb67eef51305a30b76bd7fb890603ba7d90af Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Wed, 29 Apr 2026 17:19:23 +0000
Subject: [PATCH 03/13] Add stylistic fixes addressing the linter

---
 simulstream/server/speech_processors/base_streamatt.py |  4 ++--
 .../server/speech_processors/canary_streamatt.py       | 10 ++++------
 2 files changed, 6 insertions(+), 8 deletions(-)

diff --git a/simulstream/server/speech_processors/base_streamatt.py b/simulstream/server/speech_processors/base_streamatt.py
index b0899e3..ffd99b8 100644
--- a/simulstream/server/speech_processors/base_streamatt.py
+++ b/simulstream/server/speech_processors/base_streamatt.py
@@ -62,8 +62,8 @@ class BaseStreamAtt(BaseSpeechProcessor):
              Defaults to 1.
            - **mel_hop_samples (int)**: Number of raw waveform samples per mel frame.
              Defaults to 1.
-           - **use_raw_audio_history (bool)**: Returns whether ``audio_history`` stores raw waveform samples
-             rather than processed frames. Defaults to False.
+           - **use_raw_audio_history (bool)**: Returns whether ``audio_history`` stores raw
+             waveform samples rather than processed frames. Defaults to False.
            - **text_history_max_len (int)**: The maximum length of the textual history after which
              the current content is cut. Defaults to 128.
            - **cross_attention_layer (int)**: Layer from which to extract the cross-attention from.
diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
index cec237c..4f80565 100755
--- a/simulstream/server/speech_processors/canary_streamatt.py
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -23,7 +23,7 @@
 import copy
 
 from simulstream.server.speech_processors import SAMPLE_RATE
-from simulstream.server.speech_processors.base_streamatt import BaseStreamAtt, BOW_PREFIX
+from simulstream.server.speech_processors.base_streamatt import BaseStreamAtt
 
 from nemo.collections.asr.models import ASRModel
 from nemo.collections.asr.parts.submodules.multitask_decoding import (
@@ -44,7 +44,7 @@ class CanaryStreamAtt(BaseStreamAtt):
     StreamAtt policy implementation for NVIDIA's Canary-v2 model
 
     Args:
-        config (SimpleNamespace): Configuration object.  
+        config (SimpleNamespace): Configuration object.
             Supported attributes:
             - **pnc (str)**: ``"yes"`` for punctuation/capitalisation, ``"no"`` otherwise.
               Defaults to ``"yes"``.
@@ -57,8 +57,7 @@ def __init__(self, config: SimpleNamespace):
         self.use_raw_audio_history = True
         self.mel_hop_samples = MEL_HOP_SAMPLES
         self.audio_subsampling_factor = CANARY_AUDIO_SUBSAMPLING
-        self._pnc: str = getattr(self.config, "pnc", "yes")
-        self._audio_history_max_duration: int = getattr(self.config, "audio_history_max_duration", 30)
+        self._audio_history_max_duration = getattr(self.config, "audio_history_max_duration", 30)
 
         # Build the transcription config, which is reused for every transcribe() call.
         self.transcription_cfg = MultiTaskTranscriptionConfig(
@@ -105,7 +104,6 @@ def _build_transcription_config(self):
         default_slots = copy.deepcopy(default_turns[0]["slots"])
         default_slots["source_lang"] = self.src_lang
         default_slots["target_lang"] = self.tgt_lang
-        default_slots["pnc"] = self._pnc
 
         turns = [
             {
@@ -164,4 +162,4 @@ def _generate(self, speech: np.ndarray) -> Tuple[List[str], torch.Tensor]:
         return tokens, xatt
 
     def tokens_to_string(self, tokens: List[str]) -> str:
-        return self.model.tokenizer.tokens_to_text(tokens)
\ No newline at end of file
+        return self.model.tokenizer.tokens_to_text(tokens)

From b9ec9ba15aa4409ccdbb5e9427653c181578c9ed Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Sat, 2 May 2026 13:56:08 +0000
Subject: [PATCH 04/13] Add minor fixes

---
 ...y_streamaat.yaml => canary_streamatt.yaml} |  1 +
 .../speech_processors/base_streamatt.py       | 13 +++----
 .../speech_processors/canary_streamatt.py     | 37 +++++++++++++------
 3 files changed, 32 insertions(+), 19 deletions(-)
 rename config/{canary_streamaat.yaml => canary_streamatt.yaml} (85%)

diff --git a/config/canary_streamaat.yaml b/config/canary_streamatt.yaml
similarity index 85%
rename from config/canary_streamaat.yaml
rename to config/canary_streamatt.yaml
index 861a4f7..a353714 100755
--- a/config/canary_streamaat.yaml
+++ b/config/canary_streamatt.yaml
@@ -9,5 +9,6 @@ cross_attn_layer: -2
 cutoff_frame_num: 8
 num_beams: 5
 audio_history_max_duration: 160  # Maximum length for the audio buffer, in seconds
+mel_hop_samples: 160  # Number of audio samples between adjacent mel frames
 text_history_max_len: 128
 word_level_postprocess: True  # Disable if character-level language
diff --git a/simulstream/server/speech_processors/base_streamatt.py b/simulstream/server/speech_processors/base_streamatt.py
index ffd99b8..1ec877c 100644
--- a/simulstream/server/speech_processors/base_streamatt.py
+++ b/simulstream/server/speech_processors/base_streamatt.py
@@ -81,8 +81,11 @@ def __init__(self, config: SimpleNamespace):
         text_history_cls = class_load(text_history_config.type)
         self.text_history_method = text_history_cls(text_history_config)
         self.audio_subsampling_factor = getattr(self.config, "audio_subsampling_factor", 1)
-        self.mel_hop_samples = getattr(self.config, "mel_hop_samples", 1)
+        self.mel_hop_samples = getattr(self.config, "mel_hop_samples", 160)
         self.use_raw_audio_history = getattr(self.config, "use_raw_audio_history", False)
+        self.frames_to_audio_history = self.audio_subsampling_factor
+        if self.use_raw_audio_history:
+            self.frames_to_audio_history *= self.mel_hop_samples
         self.text_history_max_len = getattr(self.config, "text_history_max_len", 128)
         self.cross_attn_layer = getattr(self.config, "cross_attention_layer", 3)
         self.cutoff_frame_num = getattr(self.config, "cutoff_frame_num", 2)
@@ -179,12 +182,8 @@ def _update_speech_history(self, discarded_text: int, cross_attn: torch.Tensor)
             # Only one token: use the unique most attended frame
             earliest_attended_idx = most_attended_idxs[0]
 
-        # Multiply by the subsampling factor to recover the original number of frames
-        frames_to_cut = earliest_attended_idx * self.audio_subsampling_factor
-
-        # If audio is stored as raw waveform, convert frames to cut into number of samples to cut
-        if self.use_raw_audio_history:
-            frames_to_cut = frames_to_cut * self.mel_hop_samples
+        # Multiply by the number of frames/samples corresponding to the audio history
+        frames_to_cut = earliest_attended_idx * self.frames_to_audio_history
 
         # Cut the unattended audio features
         self.audio_history = self.audio_history[frames_to_cut:]
diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
index 4f80565..2e85922 100755
--- a/simulstream/server/speech_processors/canary_streamatt.py
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -31,13 +31,11 @@
 )
 from nemo.collections.asr.models.aed_multitask_models import (
     MultiTaskTranscriptionConfig,
+    parse_multitask_prompt,
 )
 
 logger = logging.getLogger(__name__)
 
-MEL_HOP_SAMPLES = 160
-CANARY_AUDIO_SUBSAMPLING = 8
-
 
 class CanaryStreamAtt(BaseStreamAtt):
     """
@@ -46,19 +44,28 @@ class CanaryStreamAtt(BaseStreamAtt):
     Args:
         config (SimpleNamespace): Configuration object.
             Supported attributes:
-            - **pnc (str)**: ``"yes"`` for punctuation/capitalisation, ``"no"`` otherwise.
-              Defaults to ``"yes"``.
             - **audio_history_max_duration (int)**: Maximum audio history in seconds.
               Defaults to ``30``.
+            - **num_beams (int)**: Number of beams to use for beam search decoding.
+              Defaults to ``5``.
     """
 
     def __init__(self, config: SimpleNamespace):
         super().__init__(config)
         self.use_raw_audio_history = True
-        self.mel_hop_samples = MEL_HOP_SAMPLES
-        self.audio_subsampling_factor = CANARY_AUDIO_SUBSAMPLING
+        self.mel_hop_samples = getattr(self.config, "mel_hop_samples", 160)
+        self.audio_subsampling_factor = getattr(self.config, "audio_subsampling_factor", 8)
         self._audio_history_max_duration = getattr(self.config, "audio_history_max_duration", 30)
 
+        expected_mel_hop_samples = (
+            self.model.cfg.preprocessor.window_stride * self.model.cfg.preprocessor.sample_rate
+        )
+
+        assert self.mel_hop_samples == expected_mel_hop_samples, (
+            f"mel_hop_samples is set to {self.mel_hop_samples} in the config, but the loaded "
+            f"model's preprocessor uses {expected_mel_hop_samples} samples per mel frame"
+        )
+
         # Build the transcription config, which is reused for every transcribe() call.
         self.transcription_cfg = MultiTaskTranscriptionConfig(
             batch_size=1,
@@ -117,7 +124,12 @@ def _build_transcription_config(self):
             },
         ]
 
-        return replace(self.transcription_cfg, prompt={"turns": turns})
+        cfg_copy = copy.deepcopy(self.transcription_cfg)
+        cfg_copy.prompt = turns
+
+        logger.info(f"{self.model.tokenizer.tokens_to_text(self.text_history)=}")
+
+        return cfg_copy
 
     def _remove_eos_tokens(self, token_ids: List[int]) -> List[int]:
         """Strip leading EOS tokens that the model may prepend when a forced prefix is used."""
@@ -135,11 +147,12 @@ def _preprocess(self, waveform: np.ndarray) -> np.ndarray:
         Returns:
             np.ndarray: Accumulated raw audio history.
         """
+        waveform = waveform.astype(np.float32)
         if self.audio_history is None:
-            self.audio_history = waveform.astype(np.float32)
+            self.audio_history = waveform
         else:
             self.audio_history = np.concatenate(
-                [self.audio_history, waveform.astype(np.float32)])
+                [self.audio_history, waveform])
 
         return self.audio_history
 
@@ -153,10 +166,10 @@ def _generate(self, speech: np.ndarray) -> Tuple[List[str], torch.Tensor]:
 
         token_ids = hypothesis.y_sequence.detach().cpu().tolist()
         token_ids = self._remove_eos_tokens(token_ids)
-        tokens: List[str] = self.model.tokenizer.ids_to_tokens(token_ids)
+        tokens = self.model.tokenizer.ids_to_tokens(token_ids)
 
         xatt_raw = hypothesis.xatt_scores[self.cross_attn_layer]
-        xatt = xatt_raw.mean(dim=0).cpu()
+        xatt = xatt_raw.mean(dim=0).cpu()  # we average over heads
         xatt = self.normalize_attn(xatt)
 
         return tokens, xatt

From 056ec4e9b249c79320c8b848560a17d3ed82063e Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Sat, 2 May 2026 13:56:58 +0000
Subject: [PATCH 05/13] Fix linter issues

---
 simulstream/server/speech_processors/canary_streamatt.py | 2 --
 1 file changed, 2 deletions(-)

diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
index 2e85922..b289db1 100755
--- a/simulstream/server/speech_processors/canary_streamatt.py
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -19,7 +19,6 @@
 from types import SimpleNamespace
 from typing import List, Tuple
 
-from dataclasses import replace
 import copy
 
 from simulstream.server.speech_processors import SAMPLE_RATE
@@ -31,7 +30,6 @@
 )
 from nemo.collections.asr.models.aed_multitask_models import (
     MultiTaskTranscriptionConfig,
-    parse_multitask_prompt,
 )
 
 logger = logging.getLogger(__name__)

From 39f13805d28101fceeb884472ea9847c75dc223a Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Mon, 4 May 2026 14:07:07 +0000
Subject: [PATCH 06/13] Add minor fixes

---
 config/canary_streamatt.yaml                             | 2 ++
 simulstream/server/speech_processors/base_streamatt.py   | 2 +-
 simulstream/server/speech_processors/canary_streamatt.py | 5 -----
 3 files changed, 3 insertions(+), 6 deletions(-)

diff --git a/config/canary_streamatt.yaml b/config/canary_streamatt.yaml
index a353714..65d0976 100755
--- a/config/canary_streamatt.yaml
+++ b/config/canary_streamatt.yaml
@@ -8,7 +8,9 @@ detokenizer_type: "canary"
 cross_attn_layer: -2
 cutoff_frame_num: 8
 num_beams: 5
+audio_subsampling_factor: 8
 audio_history_max_duration: 160  # Maximum length for the audio buffer, in seconds
 mel_hop_samples: 160  # Number of audio samples between adjacent mel frames
 text_history_max_len: 128
 word_level_postprocess: True  # Disable if character-level language
+use_raw_audio_history: True
\ No newline at end of file
diff --git a/simulstream/server/speech_processors/base_streamatt.py b/simulstream/server/speech_processors/base_streamatt.py
index 1ec877c..9fb65da 100644
--- a/simulstream/server/speech_processors/base_streamatt.py
+++ b/simulstream/server/speech_processors/base_streamatt.py
@@ -61,7 +61,7 @@ class BaseStreamAtt(BaseSpeechProcessor):
            - **audio_subsampling_factor (int)**: Subsampling factor of the model, if any.
              Defaults to 1.
            - **mel_hop_samples (int)**: Number of raw waveform samples per mel frame.
-             Defaults to 1.
+             Defaults to 160, i.e. 10ms at 16kHz.
            - **use_raw_audio_history (bool)**: Returns whether ``audio_history`` stores raw
              waveform samples rather than processed frames. Defaults to False.
            - **text_history_max_len (int)**: The maximum length of the textual history after which
diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
index b289db1..bf0f387 100755
--- a/simulstream/server/speech_processors/canary_streamatt.py
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -50,9 +50,6 @@ class CanaryStreamAtt(BaseStreamAtt):
 
     def __init__(self, config: SimpleNamespace):
         super().__init__(config)
-        self.use_raw_audio_history = True
-        self.mel_hop_samples = getattr(self.config, "mel_hop_samples", 160)
-        self.audio_subsampling_factor = getattr(self.config, "audio_subsampling_factor", 8)
         self._audio_history_max_duration = getattr(self.config, "audio_history_max_duration", 30)
 
         expected_mel_hop_samples = (
@@ -125,8 +122,6 @@ def _build_transcription_config(self):
         cfg_copy = copy.deepcopy(self.transcription_cfg)
         cfg_copy.prompt = turns
 
-        logger.info(f"{self.model.tokenizer.tokens_to_text(self.text_history)=}")
-
         return cfg_copy
 
     def _remove_eos_tokens(self, token_ids: List[int]) -> List[int]:

From 3f9a6eb7fe6d60152aab6bd76241d78c2ccac054 Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Tue, 5 May 2026 20:13:15 +0000
Subject: [PATCH 07/13] Delete removing eos in the beginning

---
 .../server/speech_processors/canary_streamatt.py       | 10 ----------
 1 file changed, 10 deletions(-)

diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
index bf0f387..5363589 100755
--- a/simulstream/server/speech_processors/canary_streamatt.py
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -124,15 +124,6 @@ def _build_transcription_config(self):
 
         return cfg_copy
 
-    def _remove_eos_tokens(self, token_ids: List[int]) -> List[int]:
-        """Strip leading EOS tokens that the model may prepend when a forced prefix is used."""
-        if not token_ids:
-            return token_ids
-        pos = 0
-        while pos < len(token_ids) and token_ids[pos] == self.model.tokenizer.eos_id:
-            pos += 1
-        return token_ids[pos:]
-
     def _preprocess(self, waveform: np.ndarray) -> np.ndarray:
         """
         Append the incoming waveform chunk to the raw audio history and return it.
@@ -158,7 +149,6 @@ def _generate(self, speech: np.ndarray) -> Tuple[List[str], torch.Tensor]:
         hypothesis = output[0]
 
         token_ids = hypothesis.y_sequence.detach().cpu().tolist()
-        token_ids = self._remove_eos_tokens(token_ids)
         tokens = self.model.tokenizer.ids_to_tokens(token_ids)
 
         xatt_raw = hypothesis.xatt_scores[self.cross_attn_layer]

From 6b23ddff31b773aaf3f29aaa2618a6741124d6e3 Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Tue, 5 May 2026 21:03:22 +0000
Subject: [PATCH 08/13] Add unit test for audio trimming in update history

---
 uts/speech_processors/test_streamatt.py | 38 ++++++++++++++++++++++++-
 1 file changed, 37 insertions(+), 1 deletion(-)

diff --git a/uts/speech_processors/test_streamatt.py b/uts/speech_processors/test_streamatt.py
index 180c408..a26bd0c 100644
--- a/uts/speech_processors/test_streamatt.py
+++ b/uts/speech_processors/test_streamatt.py
@@ -14,8 +14,15 @@
 
 import unittest
 from types import SimpleNamespace
+from unittest.mock import MagicMock
+import torch
+import numpy as np
 
-from simulstream.server.speech_processors.base_streamatt import PunctuationTextHistory
+
+from simulstream.server.speech_processors.base_streamatt import (
+    BaseStreamAtt,
+    PunctuationTextHistory,
+)
 
 
 class TestPunctuationTextHistory(unittest.TestCase):
@@ -60,5 +67,34 @@ def test_no_strong_punctuation(self):
         self.assertEqual(selected_history, ['回', '到', '纽', '约', '后', '，', '我'])
 
 
+def _make_mock(text_history, audio_history, frames_to_audio_history, audio_max_len=100_000):
+    proc = MagicMock()
+    proc.text_history = text_history
+    proc.audio_history = audio_history
+    proc.frames_to_audio_history = frames_to_audio_history
+    proc.audio_max_len = audio_max_len
+    proc._cut_audio_exceeding_maxlen.side_effect = \
+        lambda: BaseStreamAtt._cut_audio_exceeding_maxlen(proc)
+    return proc
+
+
+def _cross_attn(n_text_tokens, n_audio_frames, earliest_attended_frame, discarded_text=0):
+    attn = torch.zeros(discarded_text + n_text_tokens, n_audio_frames)
+    for i in range(discarded_text, discarded_text + n_text_tokens):
+        attn[i, earliest_attended_frame] = 1.0
+    return attn
+
+
+class TestUpdateSpeechHistory(unittest.TestCase):
+    def test_trim_audio_history(self):
+        """ Test that audio history is trimmed correctly """
+        audio = np.arange(40, dtype=np.float32)
+        proc = _make_mock(["▁hello"], audio.copy(), frames_to_audio_history=4)
+        attn = _cross_attn(
+            n_text_tokens=1, n_audio_frames=10, earliest_attended_frame=2, discarded_text=1)
+        BaseStreamAtt._update_speech_history(proc, discarded_text=1, cross_attn=attn)
+        np.testing.assert_array_equal(proc.audio_history, audio[8:])
+
+
 if __name__ == "__main__":
     unittest.main()

From 52803f49a2ff83b76a1f49d156e8352a432cf8a5 Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Tue, 5 May 2026 21:10:52 +0000
Subject: [PATCH 09/13] Change the canary dependency version

---
 pyproject.toml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/pyproject.toml b/pyproject.toml
index 45a6581..3579e9b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -52,7 +52,7 @@ hf = [
 
 canary = [
     "Cython",
-    "nemo_toolkit[asr]==2.4.0",
+    "nemo_toolkit[asr]==2.8.0",
 ]
 
 vad = [

From cbc895e65df3eba23915d3bc726d30ca7e3241a0 Mon Sep 17 00:00:00 2001
From: Aziz Sharipov <101587881+azziko@users.noreply.github.com>
Date: Wed, 6 May 2026 10:36:27 +0200
Subject: [PATCH 10/13] Update
 simulstream/server/speech_processors/canary_streamatt.py

Co-authored-by: Marco Gaido <marcogaido91@gmail.com>
---
 simulstream/server/speech_processors/canary_streamatt.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
index 5363589..1a2f68f 100755
--- a/simulstream/server/speech_processors/canary_streamatt.py
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -37,7 +37,7 @@
 
 class CanaryStreamAtt(BaseStreamAtt):
     """
-    StreamAtt policy implementation for NVIDIA's Canary-v2 model
+    StreamAtt policy implementation for NVIDIA's Canary-v2 model.
 
     Args:
         config (SimpleNamespace): Configuration object.

From 59bf6f3042bf68d9b2af56b0d09d565357768757 Mon Sep 17 00:00:00 2001
From: Aziz Sharipov <101587881+azziko@users.noreply.github.com>
Date: Wed, 6 May 2026 10:37:25 +0200
Subject: [PATCH 11/13] Update uts/speech_processors/test_streamatt.py

Co-authored-by: Marco Gaido <marcogaido91@gmail.com>
---
 uts/speech_processors/test_streamatt.py | 70 +++++++++++++++++--------
 1 file changed, 49 insertions(+), 21 deletions(-)

diff --git a/uts/speech_processors/test_streamatt.py b/uts/speech_processors/test_streamatt.py
index a26bd0c..5b46f4c 100644
--- a/uts/speech_processors/test_streamatt.py
+++ b/uts/speech_processors/test_streamatt.py
@@ -67,34 +67,62 @@ def test_no_strong_punctuation(self):
         self.assertEqual(selected_history, ['回', '到', '纽', '约', '后', '，', '我'])
 
 
-def _make_mock(text_history, audio_history, frames_to_audio_history, audio_max_len=100_000):
-    proc = MagicMock()
-    proc.text_history = text_history
-    proc.audio_history = audio_history
-    proc.frames_to_audio_history = frames_to_audio_history
-    proc.audio_max_len = audio_max_len
-    proc._cut_audio_exceeding_maxlen.side_effect = \
-        lambda: BaseStreamAtt._cut_audio_exceeding_maxlen(proc)
-    return proc
+class FakeStreamAtt(BaseStreamAtt):
 
+    def _preprocess(self, waveform: np.float32) -> Union[Dict[str, torch.Tensor], torch.Tensor]:
+        raise NotImplementedError("_preprocess not implemented in FakeStreamAtt")
 
-def _cross_attn(n_text_tokens, n_audio_frames, earliest_attended_frame, discarded_text=0):
-    attn = torch.zeros(discarded_text + n_text_tokens, n_audio_frames)
-    for i in range(discarded_text, discarded_text + n_text_tokens):
-        attn[i, earliest_attended_frame] = 1.0
-    return attn
+    @classmethod
+    def load_model(cls, config: SimpleNamespace):
+        raise NotImplementedError("load_model not implemented in FakeStreamAtt")
+
+    def set_source_language(self, language: str) -> None:
+        pass
+
+    def set_target_language(self, language: str) -> None:
+        pass
+
+    def tokens_to_string(self, tokens: List[str]) -> str:
+        return " ".join(tokens)
+
+    def _generate(self, speech: torch.Tensor) -> Tuple[List[str], torch.Tensor]:
+        raise NotImplementedError("_generate not implemented in FakeStreamAtt")
+
+    @property
+    def audio_max_len(self) -> float:
+        return 10000
 
 
 class TestUpdateSpeechHistory(unittest.TestCase):
-    def test_trim_audio_history(self):
-        """ Test that audio history is trimmed correctly """
+    def _run_update_speech_history(self, use_raw_audio_history):
+        config = SimpleNamespace(
+            use_raw_audio_history=use_raw_audio_history,
+            audio_subsampling_factor=2,
+            mel_hop_samples=2,
+            text_history=SimpleNamespace(
+                type="simulstream.server.speech_processors.base_streamatt.FixedWordsTextHistory",
+            )
+
+        )
         audio = np.arange(40, dtype=np.float32)
-        proc = _make_mock(["▁hello"], audio.copy(), frames_to_audio_history=4)
-        attn = _cross_attn(
-            n_text_tokens=1, n_audio_frames=10, earliest_attended_frame=2, discarded_text=1)
-        BaseStreamAtt._update_speech_history(proc, discarded_text=1, cross_attn=attn)
-        np.testing.assert_array_equal(proc.audio_history, audio[8:])
+        proc = FakeStreamAtt(config)
+        proc.text_history = ["▁hello"]
+        proc.audio_history = audio.copy()
+
+        attn = torch.zeros(2, 10)
+        attn[1, 2] = 1.0
+
+        proc._update_speech_history(discarded_text=1, cross_attn=attn)
+        return proc.audio_history.tolist()
 
+    def test_update_speech_history_trims_audio_with_raw_audio(self):
+        audio_hist = self._run_update_speech_history(use_raw_audio_history=True)
+        # 2 audio token discarded, subsampling factor is 2, num mel hop is 2, so 2*2*2=8 samples removed 
+        self.assertListEqual(audio_hist, list(np.arange(8, 40, dtype=np.float32)))
 
+    def test_update_speech_history_trims_audio(self):
+        audio_hist = self._run_update_speech_history(use_raw_audio_history=False)
+        # 2 audio token discarded, subsampling factor is 2, so 2*2=4 samples removed 
+        self.assertListEqual(audio_hist, list(np.arange(4, 40, dtype=np.float32)))
 if __name__ == "__main__":
     unittest.main()

From 076ed3711fe348c9e5bd9a0da6ffb0c4c6d3ef92 Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Wed, 6 May 2026 08:44:20 +0000
Subject: [PATCH 12/13] Fix linter

---
 uts/speech_processors/test_streamatt.py | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/uts/speech_processors/test_streamatt.py b/uts/speech_processors/test_streamatt.py
index 5b46f4c..52cabf3 100644
--- a/uts/speech_processors/test_streamatt.py
+++ b/uts/speech_processors/test_streamatt.py
@@ -14,9 +14,9 @@
 
 import unittest
 from types import SimpleNamespace
-from unittest.mock import MagicMock
 import torch
 import numpy as np
+from typing import Dict, List, Tuple, Union
 
 
 from simulstream.server.speech_processors.base_streamatt import (
@@ -117,12 +117,15 @@ def _run_update_speech_history(self, use_raw_audio_history):
 
     def test_update_speech_history_trims_audio_with_raw_audio(self):
         audio_hist = self._run_update_speech_history(use_raw_audio_history=True)
-        # 2 audio token discarded, subsampling factor is 2, num mel hop is 2, so 2*2*2=8 samples removed 
+        # 2 audio token discarded, subsampling factor is 2,
+        # num mel hop is 2, so 2*2*2=8 samples removed
         self.assertListEqual(audio_hist, list(np.arange(8, 40, dtype=np.float32)))
 
     def test_update_speech_history_trims_audio(self):
         audio_hist = self._run_update_speech_history(use_raw_audio_history=False)
-        # 2 audio token discarded, subsampling factor is 2, so 2*2=4 samples removed 
+        # 2 audio token discarded, subsampling factor is 2, so 2*2=4 samples removed
         self.assertListEqual(audio_hist, list(np.arange(4, 40, dtype=np.float32)))
+
+
 if __name__ == "__main__":
     unittest.main()

From a1fea185f0192247fd6c3765e20e0e06ed5d7264 Mon Sep 17 00:00:00 2001
From: azziko <sharipov.wdev@gmail.com>
Date: Thu, 7 May 2026 19:53:42 +0000
Subject: [PATCH 13/13] Add minor fixes

---
 simulstream/server/speech_processors/canary_streamatt.py | 3 ---
 uts/speech_processors/test_streamatt.py                  | 1 -
 2 files changed, 4 deletions(-)

diff --git a/simulstream/server/speech_processors/canary_streamatt.py b/simulstream/server/speech_processors/canary_streamatt.py
index 1a2f68f..660bf00 100755
--- a/simulstream/server/speech_processors/canary_streamatt.py
+++ b/simulstream/server/speech_processors/canary_streamatt.py
@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License
 
-import logging
 import torch
 import numpy as np
 
@@ -32,8 +31,6 @@
     MultiTaskTranscriptionConfig,
 )
 
-logger = logging.getLogger(__name__)
-
 
 class CanaryStreamAtt(BaseStreamAtt):
     """
diff --git a/uts/speech_processors/test_streamatt.py b/uts/speech_processors/test_streamatt.py
index 52cabf3..296b30f 100644
--- a/uts/speech_processors/test_streamatt.py
+++ b/uts/speech_processors/test_streamatt.py
@@ -18,7 +18,6 @@
 import numpy as np
 from typing import Dict, List, Tuple, Union
 
-
 from simulstream.server.speech_processors.base_streamatt import (
     BaseStreamAtt,
     PunctuationTextHistory,