Automatic Speech Recognition
Transformers
Safetensors
Arabic
Moroccan Arabic
qwen3_asr
darija
fihria
moroccan-arabic
qwen3-asr
code-switching
Instructions to use fihriaLab/fihrIA-ASR-1.7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use fihriaLab/fihrIA-ASR-1.7B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="fihriaLab/fihrIA-ASR-1.7B")# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("fihriaLab/fihrIA-ASR-1.7B") model = AutoModelForMultimodalLM.from_pretrained("fihriaLab/fihrIA-ASR-1.7B", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Update model card instructions
Browse files
README.md
CHANGED
|
@@ -16,7 +16,7 @@ tags:
|
|
| 16 |
- code-switching
|
| 17 |
---
|
| 18 |
|
| 19 |
-
# fihrIA-ASR
|
| 20 |
|
| 21 |
fihrIA-ASR is fihrIA lab's open-weight ASR model family purpose-built for Moroccan Darija.
|
| 22 |
Based on Qwen3-ASR, trained using a shared 89k-token Darija tokenizer and evaluated under a
|
|
@@ -32,11 +32,29 @@ unified protocol, the family is built for robust, real-world transcription:
|
|
| 32 |
- **1.7B:** engineered as the high-accuracy ceiling and setting up a new SOTA for
|
| 33 |
open-weight Darija ASR models.
|
| 34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
## Technical & deployment details
|
| 36 |
|
| 37 |
This repository hosts the unquantized **BF16 safetensors** weights, configs, and 89k Darija
|
| 38 |
-
tokenizer
|
| 39 |
-
|
|
|
|
| 40 |
|
| 41 |
Repositories are hosted separately by size and format, sharing an identical layout:
|
| 42 |
|
|
@@ -49,14 +67,19 @@ Repositories are hosted separately by size and format, sharing an identical layo
|
|
| 49 |
|
| 50 |
**State of the art among open-weight Darija ASR models.** Both sizes beat every open-weight system we tested on held-out meaning error: the 1.7B reaches **6.1% SER** (vs 8.4% for Mistral's Voxtral Mini 4B and 11.2% for MoulSot) and achieves the best DarijaCS CER on both test sets, while running 7.1× faster than Voxtral on one H100. The 0.6B still beats Voxtral at **7.1% SER** in a laptop-sized footprint — 7.5× faster than Voxtral on H100 and 5.98× real time at 4-bit on a laptop CPU.
|
| 51 |
|
|
|
|
|
|
|
| 52 |
| | fihrIA-ASR 0.6B | fihrIA-ASR 1.7B | Mistral Voxtral Mini 4B | MoulSot 1.7B |
|
| 53 |
|---|---|---|---|---|
|
| 54 |
| Held-out mean SER ↓ | **7.1%** | **6.1%** | 8.4% | 11.2% |
|
| 55 |
-
| DarijaCS CER ↓ (Prism / Probe) | **7.98 / 9.80%** | **6.24 / 8.71%** | 6.68 / 9.69% | 10.27 / 14.93% |
|
| 56 |
| H100 throughput ↑ | **255×** real time | **243×** real time | 34× | 147× |
|
| 57 |
| Laptop CPU, Q4 ↑ | **5.98×** | — | — | 3.06× |
|
| 58 |
| Footprint, BF16 / Q4 | 1,447 / 721 MB | 3,830 / 1,642 MB | — | — |
|
| 59 |
|
|
|
|
|
|
|
|
|
|
| 60 |

|
| 61 |
|
| 62 |

|
|
@@ -65,33 +88,38 @@ Held-out numbers are the mean of four independent LLM judges over 508 clips. Vox
|
|
| 65 |
DarijaCS figures are its own published numbers; everything else is our rerun on the same
|
| 66 |
clips. Lower is better for SER and CER, higher for throughput.
|
| 67 |
|
| 68 |
-
## Contact
|
| 69 |
-
|
| 70 |
-
fihrIA is a Moroccan AI research lab working on Darija speech and language. Questions about
|
| 71 |
-
the model, the data, or a collaboration: **contact@fihria.ma**.
|
| 72 |
-
|
| 73 |
## Files
|
| 74 |
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
| File | Size (0.6B / 1.7B) | Description |
|
| 78 |
|---|---|---|
|
| 79 |
-
| `model.safetensors` |
|
| 80 |
| `config.json`, `generation_config.json`, `preprocessor_config.json` | — | Model, generation, and audio feature extractor configs. |
|
| 81 |
-
| `tokenizer.json`, `vocab.json`, `merges.txt`, `tokenizer_config.json`, `special_tokens_map.json`, `chat_template.json` | — | Shared 89k-token Darija tokenizer
|
|
|
|
| 82 |
|
| 83 |
-
##
|
| 84 |
|
| 85 |
-
|
| 86 |
-
`transformers` and `vLLM` backends):
|
| 87 |
|
| 88 |
```bash
|
| 89 |
pip install -U qwen-asr
|
| 90 |
-
# or with vLLM backend:
|
| 91 |
# pip install -U "qwen-asr[vllm]"
|
| 92 |
```
|
| 93 |
|
| 94 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
|
| 96 |
```python
|
| 97 |
import torch
|
|
@@ -108,17 +136,68 @@ model = Qwen3ASRModel.from_pretrained(
|
|
| 108 |
|
| 109 |
results = model.transcribe(
|
| 110 |
audio="sample.wav",
|
| 111 |
-
language=None,
|
| 112 |
)
|
| 113 |
|
| 114 |
print(results[0].language)
|
| 115 |
print(results[0].text)
|
| 116 |
```
|
| 117 |
|
| 118 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
|
| 120 |
```python
|
| 121 |
-
import torch
|
| 122 |
from qwen_asr import Qwen3ASRModel
|
| 123 |
|
| 124 |
model = Qwen3ASRModel.LLM(
|
|
@@ -132,11 +211,12 @@ results = model.transcribe(audio="sample.wav", language=None)
|
|
| 132 |
print(results[0].text)
|
| 133 |
```
|
| 134 |
|
|
|
|
|
|
|
| 135 |
## Prompt format
|
| 136 |
|
| 137 |
-
`qwen-asr` applies the canonical prompt automatically.
|
| 138 |
-
|
| 139 |
-
text):
|
| 140 |
|
| 141 |
```
|
| 142 |
<|im_start|>system\n<|im_end|>\n
|
|
@@ -144,62 +224,76 @@ text):
|
|
| 144 |
<|im_start|>assistant\n
|
| 145 |
```
|
| 146 |
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
Raw decoder outputs emit the detected language followed by `<asr_text>` and the transcript
|
| 150 |
-
(handled automatically by `qwen-asr`):
|
| 151 |
-
|
| 152 |
-
<div dir="rtl" lang="ar">language Darija<asr_text>التلفون ديالي طاح وتهرس ليه l'écran</div>
|
| 153 |
-
|
| 154 |
-
If you decode directly without `qwen-asr`, parse both forms (with or without `<asr_text>`):
|
| 155 |
|
| 156 |
-
|
| 157 |
-
import re
|
| 158 |
|
| 159 |
-
|
|
|
|
| 160 |
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
```
|
|
|
|
| 165 |
|
| 166 |
-
|
| 167 |
-
rather than `\S+` — otherwise the first Arabic word gets swallowed into the language field.
|
| 168 |
|
| 169 |
## Reference output
|
| 170 |
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
same pipeline with its own weights.
|
| 174 |
|
| 175 |
Ground truth:
|
| 176 |
|
| 177 |
<div dir="rtl" lang="ar">الtéléphone ديالي طاح وتهرس l'écran.</div>
|
| 178 |
|
| 179 |
-
|
| 180 |
|
| 181 |
<div dir="rtl" lang="ar">التلفون ديالي طاح وتهرس ليه l'écran</div>
|
| 182 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 183 |
The model writes the French word in Arabic script (`الtéléphone` → `التلفون`) and adds `ليه`
|
| 184 |
-
before `l'écran`.
|
| 185 |
|
| 186 |
## Checksums (MD5)
|
| 187 |
|
| 188 |
-
|
| 189 |
|
| 190 |
```
|
| 191 |
-
|
| 192 |
```
|
| 193 |
|
| 194 |
-
|
| 195 |
|
| 196 |
```
|
| 197 |
-
|
| 198 |
```
|
| 199 |
|
| 200 |
-
|
| 201 |
|
| 202 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 203 |
|
| 204 |
```bibtex
|
| 205 |
@misc{fihria2026asr,
|
|
@@ -223,5 +317,12 @@ If you use fihrIA-ASR, please cite the paper:
|
|
| 223 |
- Vocal music is transcribed as lyrics, and very noisy or instrumental clips can occasionally
|
| 224 |
produce short text. For noisy production pipelines, gating with a lightweight VAD (e.g.,
|
| 225 |
Silero VAD) is recommended.
|
| 226 |
-
-
|
| 227 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
- code-switching
|
| 17 |
---
|
| 18 |
|
| 19 |
+
# fihrIA-ASR-1.7B
|
| 20 |
|
| 21 |
fihrIA-ASR is fihrIA lab's open-weight ASR model family purpose-built for Moroccan Darija.
|
| 22 |
Based on Qwen3-ASR, trained using a shared 89k-token Darija tokenizer and evaluated under a
|
|
|
|
| 32 |
- **1.7B:** engineered as the high-accuracy ceiling and setting up a new SOTA for
|
| 33 |
open-weight Darija ASR models.
|
| 34 |
|
| 35 |
+
This repository holds the unquantized **BF16** weights of the 1.7B, the accuracy ceiling of the
|
| 36 |
+
family. It is the GPU path; for CPU deployment use the
|
| 37 |
+
[GGUF build](https://huggingface.co/fihriaLab/fihrIA-ASR-1.7B-GGUF).
|
| 38 |
+
|
| 39 |
+
## Which build do I need?
|
| 40 |
+
|
| 41 |
+
| Deployment | Build | Notes |
|
| 42 |
+
|---|---|---|
|
| 43 |
+
| CPU (laptop, edge, no GPU) | [GGUF Q4_K_M](https://huggingface.co/fihriaLab/fihrIA-ASR-1.7B-GGUF) | see the GGUF card |
|
| 44 |
+
| NVIDIA GPU | **this repo** (BF16) | `qwen-asr`, transformers backend |
|
| 45 |
+
| NVIDIA GPU, high throughput | **this repo** + vLLM | batching, `qwen-asr[vllm]` |
|
| 46 |
+
| Edge / real-time laptop CPU | [0.6B](https://huggingface.co/fihriaLab/fihrIA-ASR-0.6B) | smaller, faster, slightly higher SER |
|
| 47 |
+
|
| 48 |
+
**Memory:** 1.7B BF16 needs roughly 8 GB of VRAM (3.82 GB of weights plus KV cache and
|
| 49 |
+
activations). The 0.6B needs about 4 GB. BF16 inference on CPU runs far below real time — use
|
| 50 |
+
GGUF there.
|
| 51 |
+
|
| 52 |
## Technical & deployment details
|
| 53 |
|
| 54 |
This repository hosts the unquantized **BF16 safetensors** weights, configs, and 89k Darija
|
| 55 |
+
tokenizer. Inference runs through [`qwen-asr`](https://github.com/QwenLM/Qwen3-ASR), which
|
| 56 |
+
provides both the `transformers` and `vLLM` backends and registers the `qwen3_asr` architecture
|
| 57 |
+
with `transformers` on import.
|
| 58 |
|
| 59 |
Repositories are hosted separately by size and format, sharing an identical layout:
|
| 60 |
|
|
|
|
| 67 |
|
| 68 |
**State of the art among open-weight Darija ASR models.** Both sizes beat every open-weight system we tested on held-out meaning error: the 1.7B reaches **6.1% SER** (vs 8.4% for Mistral's Voxtral Mini 4B and 11.2% for MoulSot) and achieves the best DarijaCS CER on both test sets, while running 7.1× faster than Voxtral on one H100. The 0.6B still beats Voxtral at **7.1% SER** in a laptop-sized footprint — 7.5× faster than Voxtral on H100 and 5.98× real time at 4-bit on a laptop CPU.
|
| 69 |
|
| 70 |
+
BF16 weights (this repo):
|
| 71 |
+
|
| 72 |
| | fihrIA-ASR 0.6B | fihrIA-ASR 1.7B | Mistral Voxtral Mini 4B | MoulSot 1.7B |
|
| 73 |
|---|---|---|---|---|
|
| 74 |
| Held-out mean SER ↓ | **7.1%** | **6.1%** | 8.4% | 11.2% |
|
| 75 |
+
| DarijaCS CER ↓ (Prism / Probe, BF16) | **7.98 / 9.80%** | **6.24 / 8.71%** | 6.68 / 9.69% | 10.27 / 14.93% |
|
| 76 |
| H100 throughput ↑ | **255×** real time | **243×** real time | 34× | 147× |
|
| 77 |
| Laptop CPU, Q4 ↑ | **5.98×** | — | — | 3.06× |
|
| 78 |
| Footprint, BF16 / Q4 | 1,447 / 721 MB | 3,830 / 1,642 MB | — | — |
|
| 79 |
|
| 80 |
+
The Q4 row and footprint are measured on the GGUF builds, where the Q4 CER figures are also
|
| 81 |
+
published (0.6B: 7.98 / 9.84%; 1.7B: 6.54 / 9.23%). Use those when you deploy Q4.
|
| 82 |
+
|
| 83 |

|
| 84 |
|
| 85 |

|
|
|
|
| 88 |
DarijaCS figures are its own published numbers; everything else is our rerun on the same
|
| 89 |
clips. Lower is better for SER and CER, higher for throughput.
|
| 90 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
## Files
|
| 92 |
|
| 93 |
+
| File | Size | Description |
|
|
|
|
|
|
|
| 94 |
|---|---|---|
|
| 95 |
+
| `model.safetensors` | 3.82 GB | Full BF16 weights (audio encoder + multimodal projector + language model). |
|
| 96 |
| `config.json`, `generation_config.json`, `preprocessor_config.json` | — | Model, generation, and audio feature extractor configs. |
|
| 97 |
+
| `tokenizer.json`, `vocab.json`, `merges.txt`, `tokenizer_config.json`, `special_tokens_map.json`, `chat_template.json` | — | Shared 89k-token Darija tokenizer. |
|
| 98 |
+
| `sample.wav`, `sample.txt` | 97 KB | 3.11 s, 16 kHz mono DarijaCS Probe clip and its ground truth, for the reference check. |
|
| 99 |
|
| 100 |
+
## Install
|
| 101 |
|
| 102 |
+
**GPU:**
|
|
|
|
| 103 |
|
| 104 |
```bash
|
| 105 |
pip install -U qwen-asr
|
| 106 |
+
# or, with the vLLM backend:
|
| 107 |
# pip install -U "qwen-asr[vllm]"
|
| 108 |
```
|
| 109 |
|
| 110 |
+
**CPU-only machine** — install the CPU torch wheel first so `accelerate` does not pull unused
|
| 111 |
+
CUDA libraries:
|
| 112 |
+
|
| 113 |
+
```bash
|
| 114 |
+
pip install torch --index-url https://download.pytorch.org/whl/cpu
|
| 115 |
+
pip install -U qwen-asr
|
| 116 |
+
```
|
| 117 |
+
|
| 118 |
+
`qwen-asr` 0.0.6 pins `transformers==4.57.6`, which does not include the `qwen3_asr`
|
| 119 |
+
architecture — `qwen-asr` registers it on import. Load the model through `qwen_asr` rather than
|
| 120 |
+
with bare `transformers`.
|
| 121 |
+
|
| 122 |
+
## Quick start
|
| 123 |
|
| 124 |
```python
|
| 125 |
import torch
|
|
|
|
| 136 |
|
| 137 |
results = model.transcribe(
|
| 138 |
audio="sample.wav",
|
| 139 |
+
language=None, # auto-detect; returns "Darija" for Darija audio
|
| 140 |
)
|
| 141 |
|
| 142 |
print(results[0].language)
|
| 143 |
print(results[0].text)
|
| 144 |
```
|
| 145 |
|
| 146 |
+
`audio` accepts a local path, a URL, a base64 data URL, a `(np.ndarray, sample_rate)` tuple, or
|
| 147 |
+
a list of any of these. Audio should be 16 kHz mono; other rates are resampled.
|
| 148 |
+
|
| 149 |
+
### Reading the output
|
| 150 |
+
|
| 151 |
+
The transcript is prefixed with the detected language, and the `<asr_text>` marker is
|
| 152 |
+
present or absent depending on the runtime. Parse both forms:
|
| 153 |
+
|
| 154 |
+
```python
|
| 155 |
+
import re
|
| 156 |
+
|
| 157 |
+
ASR_RE = re.compile(r"language\s+([A-Za-z]+)\s*(?:<asr_text>)?\s*(.*)", re.DOTALL)
|
| 158 |
+
|
| 159 |
+
def parse(raw: str) -> tuple[str | None, str]:
|
| 160 |
+
"""Returns (language, transcript), with or without the <asr_text> marker."""
|
| 161 |
+
m = ASR_RE.search(raw.strip())
|
| 162 |
+
return (m.group(1), m.group(2).strip()) if m else (None, raw.strip())
|
| 163 |
+
```
|
| 164 |
+
|
| 165 |
+
The language name is ASCII and runs directly into Arabic text, so match on `[A-Za-z]+` rather
|
| 166 |
+
than `\S+` — otherwise the first Arabic word is swallowed into the language field.
|
| 167 |
+
|
| 168 |
+
## Manual `transformers` loop
|
| 169 |
+
|
| 170 |
+
For prompt control or batched decoding without `Qwen3ASRModel`:
|
| 171 |
+
|
| 172 |
+
```python
|
| 173 |
+
import torch, soundfile as sf
|
| 174 |
+
import qwen_asr # registers qwen3_asr
|
| 175 |
+
from transformers import AutoProcessor, AutoModel
|
| 176 |
+
|
| 177 |
+
mid = "fihriaLab/fihrIA-ASR-1.7B"
|
| 178 |
+
processor = AutoProcessor.from_pretrained(mid)
|
| 179 |
+
model = AutoModel.from_pretrained(mid, dtype=torch.bfloat16, device_map="cuda:0")
|
| 180 |
+
|
| 181 |
+
wav, sr = sf.read("sample.wav") # 16 kHz mono
|
| 182 |
+
prompt = ("<|im_start|>system\n<|im_end|>\n"
|
| 183 |
+
"<|im_start|>user\n<|audio_start|><|audio_pad|><|audio_end|><|im_end|>\n"
|
| 184 |
+
"<|im_start|>assistant\n")
|
| 185 |
+
|
| 186 |
+
inputs = processor(text=prompt, audio=wav, sampling_rate=sr, return_tensors="pt")
|
| 187 |
+
inputs = inputs.to(model.device).to(model.dtype)
|
| 188 |
+
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
|
| 189 |
+
|
| 190 |
+
new_tokens = out.sequences[:, inputs["input_ids"].shape[1]:]
|
| 191 |
+
raw = processor.batch_decode(new_tokens, skip_special_tokens=False)[0] # keeps <asr_text>
|
| 192 |
+
print(parse(raw))
|
| 193 |
+
```
|
| 194 |
+
|
| 195 |
+
`skip_special_tokens=True` removes the `<asr_text>` marker, which is why the language label then
|
| 196 |
+
runs straight into the transcript.
|
| 197 |
+
|
| 198 |
+
## vLLM backend
|
| 199 |
|
| 200 |
```python
|
|
|
|
| 201 |
from qwen_asr import Qwen3ASRModel
|
| 202 |
|
| 203 |
model = Qwen3ASRModel.LLM(
|
|
|
|
| 211 |
print(results[0].text)
|
| 212 |
```
|
| 213 |
|
| 214 |
+
Requires `pip install "qwen-asr[vllm]"` (pins `vllm==0.14.0`) and a CUDA GPU.
|
| 215 |
+
|
| 216 |
## Prompt format
|
| 217 |
|
| 218 |
+
`qwen-asr` applies the canonical prompt automatically. For another runtime, use an empty
|
| 219 |
+
`system` turn and an **audio-only `user` turn** (no instruction text):
|
|
|
|
| 220 |
|
| 221 |
```
|
| 222 |
<|im_start|>system\n<|im_end|>\n
|
|
|
|
| 224 |
<|im_start|>assistant\n
|
| 225 |
```
|
| 226 |
|
| 227 |
+
Use greedy decoding (`temperature=0`) for reproducible output.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 228 |
|
| 229 |
+
## Output format
|
|
|
|
| 230 |
|
| 231 |
+
The decoder emits the detected language followed by `<asr_text>` and the transcript. Whether the
|
| 232 |
+
marker survives depends on the runtime:
|
| 233 |
|
| 234 |
+
| Runtime | Output shape |
|
| 235 |
+
|---|---|
|
| 236 |
+
| `qwen-asr`, auto-detect | `language Darija…` (marker may be absent) |
|
| 237 |
+
| `llama-mtmd-cli` | `language Darija<asr_text>التلفون…` |
|
| 238 |
+
| `llama-server` chat endpoint | `language Darijaالتلفون…` (no marker) |
|
| 239 |
|
| 240 |
+
Parse both forms — see [Reading the output](#reading-the-output).
|
|
|
|
| 241 |
|
| 242 |
## Reference output
|
| 243 |
|
| 244 |
+
`sample.wav` ships with this repo: 3.11 s, 16 kHz mono WAV, DarijaCS Probe, code-switched
|
| 245 |
+
Darija–French.
|
|
|
|
| 246 |
|
| 247 |
Ground truth:
|
| 248 |
|
| 249 |
<div dir="rtl" lang="ar">الtéléphone ديالي طاح وتهرس l'écran.</div>
|
| 250 |
|
| 251 |
+
Expected transcript after stripping the `language Darija` prefix and the `<asr_text>` marker:
|
| 252 |
|
| 253 |
<div dir="rtl" lang="ar">التلفون ديالي طاح وتهرس ليه l'écran</div>
|
| 254 |
|
| 255 |
+
```python
|
| 256 |
+
# end-to-end check
|
| 257 |
+
import re, torch
|
| 258 |
+
from qwen_asr import Qwen3ASRModel
|
| 259 |
+
|
| 260 |
+
ASR_RE = re.compile(r"language\s+([A-Za-z]+)\s*(?:<asr_text>)?\s*(.*)", re.DOTALL)
|
| 261 |
+
|
| 262 |
+
model = Qwen3ASRModel.from_pretrained("fihriaLab/fihrIA-ASR-1.7B",
|
| 263 |
+
dtype=torch.bfloat16, device_map="cuda:0")
|
| 264 |
+
raw = model.transcribe(audio="sample.wav", language=None)[0].text
|
| 265 |
+
m = ASR_RE.search(raw.strip())
|
| 266 |
+
lang, text = (m.group(1), m.group(2).strip()) if m else (None, raw.strip())
|
| 267 |
+
|
| 268 |
+
assert lang == "Darija" and "طاح" in text, raw
|
| 269 |
+
print(lang, "|", text)
|
| 270 |
+
```
|
| 271 |
+
|
| 272 |
The model writes the French word in Arabic script (`الtéléphone` → `التلفون`) and adds `ليه`
|
| 273 |
+
before `l'écran`. The 0.6B runs through the same pipeline with its own weights.
|
| 274 |
|
| 275 |
## Checksums (MD5)
|
| 276 |
|
| 277 |
+
1.7B:
|
| 278 |
|
| 279 |
```
|
| 280 |
+
e9e03c692d75413fbc94479e73d8c8ae model.safetensors
|
| 281 |
```
|
| 282 |
|
| 283 |
+
0.6B:
|
| 284 |
|
| 285 |
```
|
| 286 |
+
0c49e982df1c01adda76b886643f14e8 model.safetensors
|
| 287 |
```
|
| 288 |
|
| 289 |
+
`hf download` verifies these automatically; to check manually:
|
| 290 |
|
| 291 |
+
```bash
|
| 292 |
+
hf download fihriaLab/fihrIA-ASR-1.7B --local-dir ./fihria-asr-1.7b
|
| 293 |
+
cd ./fihria-asr-1.7b && md5sum model.safetensors
|
| 294 |
+
```
|
| 295 |
+
|
| 296 |
+
## Citation
|
| 297 |
|
| 298 |
```bibtex
|
| 299 |
@misc{fihria2026asr,
|
|
|
|
| 317 |
- Vocal music is transcribed as lyrics, and very noisy or instrumental clips can occasionally
|
| 318 |
produce short text. For noisy production pipelines, gating with a lightweight VAD (e.g.,
|
| 319 |
Silero VAD) is recommended.
|
| 320 |
+
- Trained on Darija speech; performance on MSA, other Maghrebi dialects, and non-Arabic speech
|
| 321 |
+
is not characterised. Not evaluated for speaker attribution or emotion.
|
| 322 |
+
- Heavier than the 0.6B: budget ~8 GB of VRAM for BF16, and prefer the 0.6B where latency or
|
| 323 |
+
memory matter more than the last fraction of accuracy.
|
| 324 |
+
|
| 325 |
+
## Contact
|
| 326 |
+
|
| 327 |
+
fihrIA is a Moroccan AI research lab working on Darija speech and language. Questions about
|
| 328 |
+
the model, the data, or a collaboration: **contact@fihria.ma**.
|