whisperstream/stream-1-mini

Stream-1 Mini 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.

This revision adds optional custom-vocabulary hints while retaining structural dictation and numbered-list support. Use these prompts and weights together. Candidates are spelling hints, not forced substitutions. Known cleanup and entity errors remain; this revision does not establish 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	484,215,648 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-vocabulary-q4_k_m.gguf prompts.json chat_template.jinja vocabulary_contract.py --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-vocabulary-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
import sys
from pathlib import Path
from urllib.request import Request, urlopen

prompts = json.loads(Path("stream-1-mini/prompts.json").read_text(encoding="utf-8"))
sys.path.insert(0, str(Path("stream-1-mini").resolve()))
from vocabulary_contract import render_messages
vocabulary = []  # Optional relevant dictionary spellings, not the entire dictionary.
mode = "chat"  # Use "email" for Email mode.
transcript = "um let's meet on tuesday no wednesday at three pm"

payload = {
    "messages": render_messages(prompts[mode], transcript, vocabulary),
    "temperature": 0,
    "top_k": 1,
    "top_p": 1,
    "repeat_penalty": 1,
    "max_tokens": 512,
    "cache_prompt": True,
    "chat_template_kwargs": {"enable_thinking": False},
}
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 1,509,343,072 bytes (about 1.51 GB).

stream-1-mini-vocabulary-q4_k_m.gguf is the selected inference artifact. It contains 484,215,648 bytes; SHA256:

32bef61bdbc98787ac278d96a2aad120a9fc61f588f1f0a4746441e9d6d89ae7

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>. An optional JSON <VOCABULARY> block follows it; append the canonical spelling exception to the system prompt only when the block is nonempty. The supplied vocabulary_contract.py renders this exact contract (32 terms, 512 UTF-8 bytes). Render with the supplied plain ChatML template, add the assistant generation prefix and do not inject a thinking preamble or conversation history.

The full regression collection used llama.cpp b10537 on Vulkan. CPU app smoke checks are separate runtime evidence. Use the evaluated greedy contract with 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

This revision continues the published v2 model with a balanced cleanup, structural/list and vocabulary mixture. The selected terminal checkpoint used training seed 23 and 648 optimizer updates. Training data and internal transcripts are not distributed. FP32 and F16 files retain the conversion source; evaluation applies specifically to Q4_K_M with the fixed prompts, decoder and cache contract.

The selected candidate completed 3,409 registered requests, including all 1,600 historical C/D/E cleanup memberships, structure/list tests, vocabulary tests, preservation controls and actual production chunk requests. These are observed benchmark cases, with repeated memberships; they are not independent production events or untouched confirmation.

Provisional delivered cleanup counts across those 1,600 memberships are 1,216 pass, 12 warning, 202 hard failure, 53 critical failure and 117 unresolved. Unresolved judgments are not passes. Later focused feature review retained substantial vocabulary gains as well as failures:

Fresh positive-repair condition / 60	Previous v2 pass / unresolved	This revision pass / unresolved
Supplied vocabulary	7 / 1	31 / 11
Retrieved vocabulary	13 / 2	30 / 9

These purposively difficult cases do not estimate production recovery rates. Feature success requires the whole output, not just inclusion of a target name. Supplied-list negative controls can worsen even where vocabulary repair improves. Known critical regressions include deleting an independent reported account and substituting a distractor entity for a protected ordinary phrase.

Twenty separate observed-case CPU app calls completed without transport failures, timeouts or output caps. On the measured host, median request time was 0.97 s, maximum 2.54 s and peak server working set about 0.96 GiB. These are runtime smoke measurements, not quality qualification, whole-app RAM or microphone-to-insertion latency. Three of six paired repeats differed byte-for-byte; one repeat reversed an explicit host correction. Greedy settings and prefix reuse do not guarantee identical outputs across cache histories.

Fresh untouched confirmation was waived for this publication. Remaining critical errors and unresolved judgments are disclosed, not rescored as passes. No claim of general non-regression, CPU/GPU equivalence or dependable laptop tail latency is made. Verify consequential names, numbers, conditions and corrections.

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.