whisperstream/stream-1-mini

Stream-1 Mini v2 by Whisperstream

Stream-1 Mini turns English dictation transcripts into clean Chat and Email text on your own device. It is trained to remove fillers and accidental repeats, add punctuation, and resolve spoken corrections while preserving your meaning and voice. Email mode also formats dictated greetings, paragraphs and sign-offs.

Version 2 publishes the round-two candidate, checkpoint 624. It adds spoken new-line and new-paragraph cues, exclamation mark/point, explicit bullet-point cues, and improved numbered lists. Literal mentions and requests to someone else should remain dictated text. Use the v2 prompts with the v2 weights. This release trades some earlier cleanup behavior for improved structure and list support; it is not a claim of zero regressions.

Pass in text from your speech recognizer. Stream-1 Mini returns the cleaned text.

Specification	Value
Task	English dictation cleanup
Modes	Chat and Email
Base model	Qwen3-0.6B-Base
Parameters	751,632,384, with separate input embeddings and output weights
Formats	Full-precision FP32 Safetensors, F16 GGUF and Q4_K_M GGUF
GGUF download	484,216,768 bytes (about 484 MB)
Tested context	4,096 tokens, with up to 512 output tokens
License	Stream-1 Mini License, permitting commercial use with attribution

Structural dictation

The intended behavior is to convert in-place commands while preserving literal content. These examples describe the policy, not additional measured tests:

Dictation cue	Intended output
new line	One line break
new paragraph	A blank line
exclamation mark / exclamation point	!
Packing list. Bullet point socks. Bullet point charger.	A lead-in followed by two bullet items
First, passport. Second, charger. Third, jacket.	Three numbered items

Keep phrases such as "a new line of products" and "Please turn the report into bullet points" as dictated text. Spoken period, full stop, comma and question mark conversion is outside this release's structural-command scope.

Quickstart

Use Python 3, the Hugging Face CLI (hf), and llama.cpp b10537. Download the model and its prompts into a local folder:

hf download whisperstream/stream-1-mini stream-1-mini-q4_k_m.gguf prompts.json chat_template.jinja --local-dir stream-1-mini

Start the CPU server. On Windows, the executable may be named llama-server.exe.

llama-server -m stream-1-mini/stream-1-mini-q4_k_m.gguf --host 127.0.0.1 --port 8080 -c 4096 -b 256 -t 6 -np 1 -ngl 0 --chat-template-file stream-1-mini/chat_template.jinja

Save this as clean.py beside the stream-1-mini folder. It sends one dictation to the local server:

import json
from pathlib import Path
from urllib.request import Request, urlopen

prompts = json.loads(Path("stream-1-mini/prompts.json").read_text(encoding="utf-8"))
mode = "chat"  # Use "email" for Email mode.
transcript = "um let's meet on tuesday no wednesday at three pm"

payload = {
    "messages": [
        {"role": "system", "content": prompts[mode]},
        {"role": "user", "content": f"<TRANSCRIPT>{transcript}</TRANSCRIPT>"},
    ],
    "temperature": 0,
    "top_k": 1,
    "top_p": 1,
    "repeat_penalty": 1,
    "max_tokens": 512,
    "cache_prompt": True,
}
request = Request(
    "http://127.0.0.1:8080/v1/chat/completions",
    data=json.dumps(payload).encode("utf-8"),
    headers={"Content-Type": "application/json"},
)
with urlopen(request, timeout=120) as response:
    choice = json.load(response)["choices"][0]
if choice["finish_reason"] != "stop":
    raise RuntimeError("Cleanup did not finish. Keep the original transcript.")
print(choice["message"]["content"])

In another terminal, run python clean.py from that same folder. Change mode to "email" and replace transcript to try Email mode. Keep each request limited to its current transcript.

Files

The full-precision distribution contains four FP32 Safetensors shards, their index, tokenizer, configuration and chat template. It supports further modification without starting from quantized weights. Those full-precision weights are the retained source of the GGUF conversion; the evaluation below applies specifically to Q4_K_M with the stated serving setup.

stream-1-mini-f16.gguf provides the 16-bit GGUF export for llama.cpp-compatible runtimes. It contains 1,509,344,192 bytes (about 1.51 GB).

stream-1-mini-q4_k_m.gguf is the selected inference artifact. It contains 484,216,768 bytes; SHA256:

af8e5333c91eafb7b1f3dd332df741fbf0aefdee6129ba66cfca511e25a51c5d

prompts.json contains the exact Chat and Email system prompts. Keep these prompts and the supplied template together with the model. See LICENSE, LICENSE-APACHE-2.0, and NOTICE for terms and attribution.

Integration notes

Use the chosen mode's system prompt and one user message containing only the current transcript, enclosed by <TRANSCRIPT> and </TRANSCRIPT>. Render with the supplied plain ChatML template, add the assistant generation prefix and do not inject a thinking preamble or conversation history.

The evaluated configuration used llama.cpp b10537 on CPU, six threads, 4,096-token context, batch size 256, one slot and one request at a time. Decode greedily: temperature 0, top-k 1, top-p 1, repeat penalty 1, and at most 512 output tokens. Stop at the end-of-turn token. The training/export stop IDs are 151643 and 151645; the GGUF end-of-turn ID is 151645.

Enable prefix caching and keep Chat and Email sessions separate. Reuse a prompt prefix without adding earlier dictations to the current request. Reject incomplete or invalid generations and retain the original transcript when cleanup fails. Users should be able to inspect and recover the original text.

Training and evaluation

V2 continues the v1 E full13 model through structural-cue training and a numbered-list follow-up. The published round-two checkpoint used a 9,984-position weighted continuation mixture over two passes and 624 optimizer updates, with original-E replay and independently reviewed synthetic command/literal contrasts. Training data and internal transcripts are not included.

The Q4_K_M candidate was evaluated on four full historical cleanup cohorts: 600 C/D development, 200 C/D confirmation, 600 E development and 200 E confirmation requests. These contain 1,205 distinct source/mode combinations; overlapping cases with different request seeds remain separate observations. They are observed regression evidence, not untouched selection tests.

Configuration / 1,600	Raw acceptable	Delivered acceptable	Delivered critical	Delivered uncertain	Delivered preserved
V1 selected E	1,260	1,257	62	34	1,496
V2 round two	1,250	1,252	67	37	1,481

Acceptable means pass plus minor warning; uncertainty remains separate. Delivered results use pinned Whisperstream postprocessing, including real chunk inference for long inputs, and are not raw model accuracy. Against v1, v2 introduces 32 critical delivered failures and fixes 26; net totals do not erase individual regressions. Critical failures can include changed numbers, invented units or lost instructions. The full cleanup comparison used CPU inference with the configuration above. It does not establish production error rates or full-suite Vulkan equivalence.

On a separate fresh structural development set, v2 had 135/144 acceptable outputs on both CPU and Vulkan. Primary-feature successes were:

Feature	CPU	Vulkan
New line	10/12	9/12
New paragraph	12/12	12/12
Exclamation	11/12	12/12
Explicit bullets	12/12	12/12
Numbered lists	21/24	21/24
Primary literal controls preserved	72/72	72/72

Feature success does not imply that every other detail in an output is correct. A separate eight-case preservation extension remains only 3/8 acceptable, including failures to preserve literally discussed hesitation words. New sealed structural and numbered confirmation sets were not opened for this publication. No v2 startup, tail-latency or whole-app reliability claim is made here.

Limitations

The model can change meaning, retain an incorrect number after a correction, or infer the wrong numeric scale. Review names, numbers, restrictions and other consequential details before relying on cleaned text.

This release has not established support for code, arbitrary style controls, multilingual cleanup or general question answering. Instructions within a transcript should remain dictated content, except the supported in-place structural dictation cues described above. The model has not been validated for medical, legal or other high-stakes decisions. Smaller quantizations and changed prompts may behave differently.

License

Personal and commercial use, modification and redistribution are permitted under the Stream-1 Mini License, which combines Apache-2.0 terms with an identification requirement for Stream-1 Mini by Whisperstream. This is a custom license, not unmodified Apache-2.0. See the license for permitted attribution locations and treatment of modified versions.