0xShug0 · 0xShug0 · Sep 10, 2026 · Sep 9, 2026 · Sep 9, 2026 · Sep 10, 2026
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -968,6 +968,7 @@ audiocpp_add_model(vibevoice_asr
         engine/models/vibevoice_asr/loader.h
     LOADERS
         engine::models::vibevoice_asr::make_vibevoice_asr_loader
+        engine::models::vibevoice_asr::make_vibevoice_asr_streaming_loader
 )
 
 audiocpp_add_model(voxtral_realtime

diff --git a/docs/asr.md b/docs/asr.md
@@ -12,7 +12,8 @@
 | Nemotron ASR | `nemotron_asr` | offline, streaming | [Nemotron ASR](#nemotron-asr) |
 | Parakeet-TDT | `parakeet_tdt` | offline, streaming | [Parakeet-TDT](#parakeet-tdt) |
 | SenseVoice-Small | `sense_asr` | offline, streaming | [SenseVoice-Small](#sensevoice-small) |
-| VibeVoice ASR | `vibevoice_asr` | offline | [VibeVoice ASR](#vibevoice-asr) |
+| VibeVoice ASR | `vibevoice_asr` | offline | [VibeVoice ASR](models/vibevoice_asr.md#vibevoice-asr) |
+| VibeVoice ASR Streaming 7B | `vibevoice_asr_streaming` | offline, streaming | [VibeVoice ASR Streaming 7B](models/vibevoice_asr.md#vibevoice-asr-streaming-7b) |
 | Voxtral Realtime | `voxtral_realtime` | offline, streaming | [Voxtral Realtime](#voxtral-realtime) |
 
 This page covers ASR models. Detailed Qwen3 ASR and forced-alignment notes live in [Qwen3 models](models/qwen3.md).
@@ -327,68 +328,13 @@ chunking, server usage, and validation notes.
 
 ## VibeVoice ASR
 
-VibeVoice ASR is an offline ASR model with greedy, sampling, and beam-search decode paths. It can return transcription text and structured segment/speaker-turn output when the model produces timestamps.
+VibeVoice ASR covers the original offline ASR family and the new Streaming 7B
+family. The streaming family has its own dedicated GGUF repo and supports both
+offline and live streaming transcription. The Streaming 7B model can emit
+speaker-attributed text, but it does not produce timestamped segments.
 
-A fully quantized port of the same model — INT8 activations through the encoder,
-ternary BitNet weights in the decoder — lives under community models as
-`vibeasr`: see [VibeASR](community_models/vibeasr.md). It is not a separate
-model, only a CPU-only alternative numeric pipeline for the same weights.
-
-| Field | Value |
-|---|---|
-| Family | `vibevoice_asr` |
-| Model directory | `models/VibeVoice-ASR` |
-| Task | `asr` |
-| Modes | `offline` |
-| Required tokenizer files | `tokenizer.json`, `tokenizer_config.json`, `vocab.json`, and `merges.txt` in the model directory |
-| Output | Transcription text; optional segments through `--segments-out`; optional speaker turns through `--turns-out` |
-| Streaming | Not supported |
-| Timestamps | Segment and speaker-turn timestamps when produced |
-
-```bash
-audiocpp_cli --task asr --family vibevoice_asr --model models/VibeVoice-ASR-GGUF/vibevoice-asr-q8_0.gguf --backend cuda --audio assets/resources/sample_16k.wav --text-out transcript.txt
-```
-
-VibeVoice-ASR also accepts a standalone audio.cpp-native GGUF. Pass the shard
-index to merge all eight safetensors files while converting:
-
-```powershell
-audiocpp_gguf.exe --input models\VibeVoice-ASR\model.safetensors.index.json --output models\VibeVoice-ASR-Q8_0\model.gguf --type q8_0
-```
-
-Configuration and tokenizer assets are embedded by default, so the output
-directory may contain only `model.gguf`.
-
-Structured output:
-
-```bash
-audiocpp_cli --task asr --family vibevoice_asr --model models/VibeVoice-ASR-GGUF/vibevoice-asr-q8_0.gguf --backend cuda --audio meeting.wav --text "The recording is a meeting conversation." --text-out transcript.txt --segments-out segments.json --turns-out turns.json
-```
-
-With VAD chunking, provide the bundled Silero VAD model:
-
-```bash
-audiocpp_cli --task asr --family vibevoice_asr --model models/VibeVoice-ASR-GGUF/vibevoice-asr-q8_0.gguf --backend cuda --audio assets/resources/sample_16k.wav --audio-chunk-mode vad --session-option vibevoice_asr.vad_model_path=assets/framework/models/silero_vad --text-out transcript.txt
-```
-
-| Option | Values | Default | Meaning |
-|---|---|---:|---|
-| `--audio` | WAV path | required | Speech input. |
-| `--text` | text | empty string | Context prompt for the ASR request. |
-| `--language` | language code | `auto` | ASR language label. |
-| `--max-tokens` | integer | model default | Maximum generated transcript tokens. |
-| `--temperature` | float | model default | Sampling temperature; `0` uses deterministic decoding. |
-| `--top-p` | float | model default | Nucleus sampling probability. |
-| `--top-k` | integer | model default | Top-k sampling limit; `0` disables top-k filtering. |
-| `--num-beams` | integer | `1` | Beam count for deterministic beam search. |
-| `--repetition-penalty` | float | model default | Generation repetition penalty. |
-| `--seed` | integer | random if omitted | Sampling seed. |
-| `--audio-chunk-mode` | `auto`, `fixed`, `vad`, `none` | `auto` | Long-audio chunking mode. `auto` uses fixed chunks. |
-| `--audio-chunk-seconds` | float seconds | `1200` | Fixed audio chunk duration. |
-| `--text-out` | TXT path | not set | Transcript output. The transcript is also printed to stdout. |
-| `--segments-out` | JSON path | not set | Write structured ASR segments when produced. |
-| `--turns-out` | JSON path | not set | Write speaker turns when produced. |
-| `--session-option vibevoice_asr.vad_model_path=<path>` | model directory | `assets/framework/models/silero_vad` | Internal VAD model used by `--audio-chunk-mode vad`. |
+See [VibeVoice ASR models](models/vibevoice_asr.md) for package IDs, conversion
+notes, CLI examples, live server configuration, and request options.
 
 ## Voxtral Realtime
 

diff --git a/docs/gguf.md b/docs/gguf.md
@@ -106,6 +106,7 @@ Status labels:
 | `vevo2` | Done | Pass | Pass | Pass (drift) | No (mixed route drift; speech ASR match) |
 | `vibevoice` | Done | Pass | --- | Pass | Pass (drift) |
 | `vibevoice_asr` | Done | Pass | --- | Pass | Pass |
+| `vibevoice_asr_streaming` | Done | --- | --- | Pass | Pass |
 | `voxcpm2` | Done | Pass | Pass | Pass (ASR match, drift) | Pass (ASR match, drift) |
 | `voxtral_realtime` | Done | Pass | --- | Pass | Pass |
 
@@ -115,6 +116,7 @@ Additional lower-bit checks:
 |---|---|---|
 | `meanvc2` | `q4_k` | Pass |
 | `personaplex` | `q4_k` | Pass |
+| `vibevoice_asr_streaming` | `q4_k` | Pass (quick CUDA check; transcript stays usable and matches the BF16 wording class) |
 | `voxtral_realtime` | `q4_k` | Pass (quick CUDA check; transcripts match Q8 except one capitalization-only difference) |
 
 Q8 packaging notes:
@@ -145,6 +147,9 @@ Q8 packaging notes:
 - `voxtral_realtime` also has a tested `q4_k` package. In a quick CUDA path
   check it was smaller and faster than Q8_0, while transcript output matched
   Q8_0 except for one capitalization-only difference.
+- `vibevoice_asr_streaming` also has a tested `q4_k` package. In a quick CUDA
+  check, BF16 and Q4_K produced the same transcript wording on the validation
+  clip; Q8_0 produced the same sentence with minor wording drift.
 
 ## Build The Converter
 

diff --git a/docs/models/vibevoice_asr.md b/docs/models/vibevoice_asr.md
@@ -0,0 +1,221 @@
+# VibeVoice ASR Models
+
+audio.cpp supports two VibeVoice ASR families:
+
+| Model | Family | Mode(s) | Recommended package |
+|---|---|---|---|
+| VibeVoice ASR | `vibevoice_asr` | offline | `vibevoice_asr_q8_0` |
+| VibeVoice ASR Streaming 7B | `vibevoice_asr_streaming` | offline, streaming | `vibevoice_asr_streaming_7b_q8_0` |
+
+## VibeVoice ASR
+
+VibeVoice ASR is an offline ASR model with greedy, sampling, and beam-search
+decode paths. It can return transcription text and structured segment or
+speaker-turn output when the model produces timestamps.
+
+A fully quantized port of the same model — INT8 activations through the encoder,
+ternary BitNet weights in the decoder — lives under community models as
+`vibeasr`: see [VibeASR](../community_models/vibeasr.md). It is not a separate
+model, only a CPU-only alternative numeric pipeline for the same weights.
+
+| Field | Value |
+|---|---|
+| Family | `vibevoice_asr` |
+| Model directory | `models/VibeVoice-ASR-GGUF` |
+| Task | `asr` |
+| Modes | `offline` |
+| Required tokenizer files | Embedded in the standalone GGUF |
+| Output | Transcription text; optional segments through `--segments-out`; optional speaker turns through `--turns-out` |
+| Streaming | Not supported |
+| Timestamps | Segment and speaker-turn timestamps when produced |
+
+Install:
+
+```bash
+python3 tools/model_manager_v2.py install vibevoice_asr_q8_0
+```
+
+Offline CLI:
+
+```bash
+audiocpp_cli --task asr \
+  --family vibevoice_asr \
+  --model models/VibeVoice-ASR-GGUF/vibevoice-asr-q8_0.gguf \
+  --backend cuda \
+  --audio assets/resources/sample_16k.wav \
+  --text-out transcript.txt \
+  --metrics \
+  --log
+```
+
+Structured output:
+
+```bash
+audiocpp_cli --task asr \
+  --family vibevoice_asr \
+  --model models/VibeVoice-ASR-GGUF/vibevoice-asr-q8_0.gguf \
+  --backend cuda \
+  --audio meeting.wav \
+  --text "The recording is a meeting conversation." \
+  --text-out transcript.txt \
+  --segments-out segments.json \
+  --turns-out turns.json \
+  --metrics \
+  --log
+```
+
+With VAD chunking, provide the bundled Silero VAD model:
+
+```bash
+audiocpp_cli --task asr \
+  --family vibevoice_asr \
+  --model models/VibeVoice-ASR-GGUF/vibevoice-asr-q8_0.gguf \
+  --backend cuda \
+  --audio assets/resources/sample_16k.wav \
+  --audio-chunk-mode vad \
+  --session-option vibevoice_asr.vad_model_path=assets/framework/models/silero_vad \
+  --text-out transcript.txt \
+  --metrics \
+  --log
+```
+
+Convert from safetensors:
+
+```powershell
+audiocpp_gguf.exe --input models\VibeVoice-ASR\model.safetensors.index.json --output models\VibeVoice-ASR-Q8_0\model.gguf --type q8_0
+```
+
+Configuration and tokenizer assets are embedded by default, so the output
+directory may contain only `model.gguf`.
+
+Options:
+
+| Option | Values | Default | Meaning |
+|---|---|---:|---|
+| `--audio` | WAV path | required | Speech input. |
+| `--text` | text | empty string | Context prompt for the ASR request. |
+| `--language` | language code | `auto` | ASR language label. |
+| `--max-tokens` | integer | model default | Maximum generated transcript tokens. |
+| `--temperature` | float | model default | Sampling temperature; `0` uses deterministic decoding. |
+| `--top-p` | float | model default | Nucleus sampling probability. |
+| `--top-k` | integer | model default | Top-k sampling limit; `0` disables top-k filtering. |
+| `--num-beams` | integer | `1` | Beam count for deterministic beam search. |
+| `--repetition-penalty` | float | model default | Generation repetition penalty. |
+| `--seed` | integer | random if omitted | Sampling seed. |
+| `--audio-chunk-mode` | `auto`, `fixed`, `vad`, `none` | `auto` | Long-audio chunking mode. `auto` uses fixed chunks. |
+| `--audio-chunk-seconds` | float seconds | `1200` | Fixed audio chunk duration. |
+| `--text-out` | TXT path | not set | Transcript output. The transcript is also printed to stdout. |
+| `--segments-out` | JSON path | not set | Write structured ASR segments when produced. |
+| `--turns-out` | JSON path | not set | Write speaker turns when produced. |
+| `--session-option vibevoice_asr.vad_model_path=<path>` | model directory | `assets/framework/models/silero_vad` | Internal VAD model used by `--audio-chunk-mode vad`. |
+
+## VibeVoice ASR Streaming 7B
+
+VibeVoice ASR Streaming 7B is the streaming VibeVoice ASR model. It keeps a
+persistent decoder state and can emit speaker-attributed transcript deltas as
+audio arrives. It can also run in offline mode through the same family.
+
+The upstream streaming model emits speaker-attributed text such as
+`Speaker 0: ...`, but it does not emit timestamped segments. In audio.cpp,
+`--turns-out` can expose those speaker/text turns when the model emits speaker
+labels; `--segments-out` is not populated by this streaming model. Use the
+non-streaming `vibevoice_asr` family when timestamped segment output is needed.
+
+> [!WARNING]
+> Speaker-turn segmentation can differ from the offline VibeVoice ASR family.
+> In local validation, the streaming 7B model merged adjacent speech into fewer
+> speaker turns than the offline model; the same behavior was reproduced with
+> the official Python streaming reference.
+
+The recommended audio.cpp package is the Q8_0 GGUF in the dedicated model repo:
+<https://huggingface.co/audio-cpp/VibeVoice-ASR-Streaming-7B-GGUF>. BF16 and
+Q4_K GGUF variants are also available in the same repo.
+
+Upstream lists ten supported language codes for this streaming model: `en`,
+`zh`, `es`, `pt`, `de`, `ja`, `ko`, `fr`, `ru`, and `it`.
+
+| Field | Value |
+|---|---|
+| Family | `vibevoice_asr_streaming` |
+| Model package | `vibevoice_asr_streaming_7b_q8_0` |
+| Model directory | `models/VibeVoice-ASR-Streaming-7B-GGUF` |
+| Task | `asr` |
+| Modes | `offline`, `streaming` |
+| Required tokenizer files | Embedded in the standalone GGUF |
+| Output | Transcript text; speaker/text turns when produced by the model |
+| Timestamps | Not produced by the streaming model |
+| Streaming | Live audio chunks over the `/v1/audio/transcriptions/live` endpoint |
+
+Install:
+
+```bash
+python3 tools/model_manager_v2.py install vibevoice_asr_streaming_7b_q8_0
+```
+
+Offline CLI:
+
+```bash
+audiocpp_cli --task asr \
+  --family vibevoice_asr_streaming \
+  --model models/VibeVoice-ASR-Streaming-7B-GGUF/vibevoice-asr-streaming-7b-q8_0.gguf \
+  --backend cuda \
+  --threads 8 \
+  --audio assets/resources/sample_16k.wav \
+  --text-out transcript.txt \
+  --turns-out speaker_turns.json \
+  --metrics \
+  --log
+```
+
+Server config for live streaming:
+
+```json
+{
+  "host": "127.0.0.1",
+  "port": 8080,
+  "backend": "cuda",
+  "threads": 8,
+  "models": [
+    {
+      "id": "vibevoice-streaming-7b",
+      "family": "vibevoice_asr_streaming",
+      "path": "models/VibeVoice-ASR-Streaming-7B-GGUF/vibevoice-asr-streaming-7b-q8_0.gguf",
+      "task": "asr",
+      "mode": "streaming"
+    }
+  ]
+}
+```
+
+Start the server:
+
+```bash
+audiocpp_server --config server.json --log
+```
+
+Live streaming request with 16 kHz mono signed 16-bit PCM:
+
+```bash
+ffmpeg -hide_banner -loglevel error -i input.wav -f s16le -ac 1 -ar 16000 - \
+  | curl -N -X POST \
+      -H 'Content-Type: application/octet-stream' \
+      -H 'Transfer-Encoding: chunked' \
+      -H 'Expect:' \
+      -T - \
+      'http://127.0.0.1:8080/v1/audio/transcriptions/live?model=vibevoice-streaming-7b&sample_rate=16000&channels=1&sample_format=s16le'
+```
+
+Common request options:
+
+| Option | Values | Default | Meaning |
+|---|---|---:|---|
+| `--language` | language label | `auto` | ASR language label. |
+| `--request-option context=<text>` | text | empty string | Extra context or hotwords injected into the streaming prompt. |
+| `--max-tokens` | integer | `256` | Maximum generated transcript tokens per chunk. |
+| `--temperature` | float | `0` | Sampling temperature; `0` uses deterministic decoding. |
+| `--top-p` | float | `1` | Nucleus sampling probability. |
+| `--top-k` | integer | `0` | Top-k sampling limit; `0` disables top-k filtering. |
+| `--num-beams` | integer | `1` | Beam count for deterministic beam search. |
+| `--repetition-penalty` | float | `1` | Generation repetition penalty. |
+| `--audio-chunk-mode` | `auto`, `fixed`, `vad`, `none` | `auto` | Offline audio chunking mode. |
+| `--audio-chunk-seconds` | float seconds | `1200` | Offline chunk duration for fixed and VAD chunking. |
diff --git a/include/engine/models/vibevoice_asr/assets.h b/include/engine/models/vibevoice_asr/assets.h
@@ -87,13 +87,16 @@ struct VibeVoiceAudioProcessorConfig {
 
 struct VibeVoiceProcessorConfig {
     int64_t speech_tok_compress_ratio = 3200;
+    int64_t chunk_frames = 0;
+    int64_t lookahead_frames = 0;
     bool db_normalize = true;
     std::string language_model_pretrained_name = "Qwen/Qwen2.5-1.5B";
     VibeVoiceAudioProcessorConfig audio_processor;
 };
 
 struct VibeVoiceASRAssets {
     assets::ResourceBundle resources;
+    std::string family = "vibevoice_asr";
     VibeVoiceConfig config;
     VibeVoiceProcessorConfig processor;
     std::shared_ptr<const assets::TensorSource> model_weights;
@@ -104,6 +107,8 @@ struct VibeVoiceASRAssets {
     bool fine_tune_applied = false;
 };
 
-std::shared_ptr<const VibeVoiceASRAssets> load_vibevoice_asr_assets(const std::filesystem::path & model_path);
+std::shared_ptr<const VibeVoiceASRAssets> load_vibevoice_asr_assets(
+    const std::filesystem::path & model_path,
+    const std::string & family = "vibevoice_asr");
 
 }  // namespace engine::models::vibevoice_asr
diff --git a/include/engine/models/vibevoice_asr/connector.h b/include/engine/models/vibevoice_asr/connector.h
@@ -80,6 +80,7 @@ class VibeVoiceConnectorWeightsRuntime final {
         int64_t input_dim) const;
     std::vector<VibeVoiceConnectorOutput> project_semantic_batch(
         const std::vector<VibeVoiceTokenizerLatents> & features) const;
+    void release_cached_graphs() const;
 
 private:
     std::shared_ptr<const VibeVoiceASRAssets> assets_;

diff --git a/include/engine/models/vibevoice_asr/loader.h b/include/engine/models/vibevoice_asr/loader.h
@@ -4,6 +4,7 @@
 #include "engine/models/vibevoice_asr/assets.h"
 
 #include <memory>
+#include <string>
 
 namespace engine::models::vibevoice_asr {
 
@@ -26,7 +27,10 @@ class VibeVoiceASRLoadedModel final : public runtime::ILoadedVoiceModel {
     std::shared_ptr<const VibeVoiceASRAssets> assets_;
 };
 
-std::unique_ptr<VibeVoiceASRLoadedModel> load_vibevoice_asr_model(const std::filesystem::path & model_path);
+std::unique_ptr<VibeVoiceASRLoadedModel> load_vibevoice_asr_model(
+    const std::filesystem::path & model_path,
+    const std::string & family = "vibevoice_asr");
 std::shared_ptr<runtime::IVoiceModelLoader> make_vibevoice_asr_loader();
+std::shared_ptr<runtime::IVoiceModelLoader> make_vibevoice_asr_streaming_loader();
 
 }  // namespace engine::models::vibevoice_asr